<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Kunal Vohra</title>
    <description>The latest articles on DEV Community by Kunal Vohra (@kunalvohra).</description>
    <link>https://dev.to/kunalvohra</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F191502%2Fc6851beb-8a02-458a-84da-59794b9d9d8a.JPG</url>
      <title>DEV Community: Kunal Vohra</title>
      <link>https://dev.to/kunalvohra</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/kunalvohra"/>
    <language>en</language>
    <item>
      <title>Node.js Security Release (November 2026): How to Fix the 16 CVEs in 22.23.5, 24.22.3 and 26.13.1, Step by Step</title>
      <dc:creator>Kunal Vohra</dc:creator>
      <pubDate>Thu, 12 Nov 2026 12:00:00 +0000</pubDate>
      <link>https://dev.to/kunalvohra/nodejs-security-release-november-2026-how-to-fix-the-16-cves-in-22235-24223-and-26131-1mph</link>
      <guid>https://dev.to/kunalvohra/nodejs-security-release-november-2026-how-to-fix-the-16-cves-in-22235-24223-and-26131-1mph</guid>
      <description>&lt;p&gt;On Wednesday 11 November 2026, the Node.js project released security fixes for 16 vulnerabilities across every supported release line: 7 rated high, 6 medium and 3 low. Most of the serious ones are in HTTP/2 and TLS handling, and the worst outcomes are an unauthenticated remote crash, a use-after-free that can send another connection's heap memory to the client, and ways around client-certificate (mTLS) checks. The fixed versions are &lt;strong&gt;Node.js 22.23.5, 24.22.3 and 26.13.1&lt;/strong&gt;. If your Node process speaks HTTP/2 or terminates TLS itself, upgrade today. If it sits behind Nginx or a load balancer that terminates TLS, you have more breathing room, but you should still upgrade this week, because some of the fixes are on the client side. And if you are still on Node 20, there is no fix coming for you at all.&lt;/p&gt;

&lt;p&gt;I'm Kunal Vohra, a fractional CTO and a security researcher and cybersecurity advisor. Node.js runs a huge share of startup backends: API servers, Next.js apps, workers, CLIs, CI jobs, serverless functions. This is the second Node.js security release in three weeks (the October one fixed the Brotli bug some outlets call "Node2Shell"), so here is what changed, how to tell whether it matters for your stack, and a patch routine you can run in under an hour.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the November 2026 Node.js release fixes
&lt;/h2&gt;

&lt;p&gt;The official advisory, "Wednesday, November 11, 2026 Security Releases" on nodejs.org, lists 16 CVEs. Node.js did not publish CVSS scores in the advisory, only high, medium and low labels, and it does not say that any of these are being exploited.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The seven high-severity issues:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;CVE-2026-123863, HTTP/2 remote abort.&lt;/strong&gt; An unauthenticated client can trip an internal assertion when an HTTP/2 session is torn down and crash the process. Node.js describes it as a variant of the 2024 bug CVE-2024-27983. Affects 24.x and 26.x.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;CVE-2026-123864, HTTPS re-entrancy in the HTTP parser.&lt;/strong&gt; A zero-length TLS write can re-enter llhttp, causing a use-after-free and an out-of-bounds heap read that can leak memory into a response body. Plain &lt;code&gt;node:http&lt;/code&gt; without TLS is not affected. Affects 22.x, 24.x and 26.x.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;CVE-2026-123865, HTTP/2 sessions that never clean up.&lt;/strong&gt; A protocol error during connection shutdown leaks memory and file descriptors permanently. Node.js calls it an incomplete fix of an earlier CVE. Affects all three lines.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;CVE-2026-123866, freed memory sent to the peer.&lt;/strong&gt; During HTTP/2 stream teardown, a use-after-free can send freed memory, including data from other connections, back to the client. Affects all three lines.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;CVE-2026-126051, TLS server crash via deferred handshake callbacks.&lt;/strong&gt; Servers using &lt;code&gt;SNICallback&lt;/code&gt;, OCSP or session resumption callbacks can be crashed by an uncaught exception. Affects 24.x and 26.x.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;CVE-2026-123872, TLS 1.2 renegotiation changes the server name.&lt;/strong&gt; A client can change the SNI server name during renegotiation and bypass client-certificate authentication. Affects all three lines.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;CVE-2026-123874, HTTP/2 use-after-free in &lt;code&gt;respondWithFile()&lt;/code&gt;.&lt;/strong&gt; Only reachable if your app serves files with &lt;code&gt;respondWithFile()&lt;/code&gt; or &lt;code&gt;respondWithFD()&lt;/code&gt;. Affects 24.x and 26.x.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;The medium ones&lt;/strong&gt; include a one-byte heap overflow in &lt;code&gt;session.altsvc()&lt;/code&gt; (CVE-2026-123828), an mTLS fail-open when per-host SNI contexts only override the CA (CVE-2026-123867), unbounded HTTP/2 memory that &lt;code&gt;maxSessionMemory&lt;/code&gt; does not catch (CVE-2026-123868), proxy credentials leaking into &lt;code&gt;ERR_PROXY_INVALID_CONFIG&lt;/code&gt; error messages (CVE-2026-123871), a certificate validation bypass in &lt;code&gt;https.Agent&lt;/code&gt; through TLS session reuse across different policies (CVE-2026-123877), and a server crash when someone cancels a &lt;code&gt;respondWithFile()&lt;/code&gt; download (CVE-2026-123875).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The low ones&lt;/strong&gt; are in the experimental Permission Model and zlib: a recursive &lt;code&gt;fs.watch&lt;/code&gt; that leaks the names of denied files on macOS and Windows (CVE-2026-123829), an uncatchable assertion in zlib (CVE-2026-123862), and a wildcard file-system over-grant. One small inconsistency to know about: the advisory lists that last one as CVE-2026-106071, while the 24.22.3 release notes list CVE-2026-123873 ("permission: check heap snapshot destination") instead. It does not change what you need to do.&lt;/p&gt;

&lt;p&gt;The release also updates bundled dependencies: llhttp on all three lines, undici on 24.x and 26.x, and SQLite on 22.x and 24.x. The 24.22.3 release notes list llhttp 9.4.4, undici 7.30.2, SQLite 3.54.1 and Brotli 1.2.1.&lt;/p&gt;

&lt;p&gt;A detail worth noticing: five of the sixteen findings are credited to Chris Rohlf "in collaboration with Claude and Anthropic Research", and one to Trail of Bits (OpenAI). The October release credited the Brotli bug to OpenAI Codex Security. AI-assisted bug hunting is now feeding the patch queue of the runtime most startups depend on, which is the same pattern I wrote about in &lt;a href="https://kunalvohra.com/blog/openai-948-linux-root-exploits" rel="noopener noreferrer"&gt;OpenAI's 948 Linux root exploits&lt;/a&gt;. Expect more frequent releases, not fewer.&lt;/p&gt;

&lt;h2&gt;
  
  
  Is this you? A five-minute check
&lt;/h2&gt;

&lt;p&gt;Severity labels are a starting point. What matters is how your Node processes are exposed. Work through these in order.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Which versions are you running?&lt;/strong&gt; Anything below 22.23.5, 24.22.3 or 26.13.1 on its line is affected. Check every place Node runs, not just your laptop:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# On a server or in CI&lt;/span&gt;
node --version
node -p &lt;span class="s2"&gt;"process.versions.llhttp"&lt;/span&gt;

&lt;span class="c"&gt;# Inside a container image you deploy&lt;/span&gt;
docker run --rm your-registry/your-api:latest node --version

&lt;span class="c"&gt;# Find the base images in your repos&lt;/span&gt;
grep -rn &lt;span class="s2"&gt;"FROM node"&lt;/span&gt; --include=Dockerfile* .

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;2. Does Node terminate TLS or speak HTTP/2 itself?&lt;/strong&gt; Most of the high-severity bugs live here. If your Node app listens directly on 443, creates an &lt;code&gt;https&lt;/code&gt; or &lt;code&gt;http2&lt;/code&gt; server, or uses mTLS, you are in the most exposed group. Search your code:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;grep -rnE &lt;span class="s2"&gt;"http2\.create(Secure)?Server|https\.createServer|tls\.createServer|SNICallback|requestCert|respondWith(File|FD)|altsvc"&lt;/span&gt; --include=*.&lt;span class="o"&gt;{&lt;/span&gt;js,ts,mjs,cjs&lt;span class="o"&gt;}&lt;/span&gt; src/ server/ 2&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;/dev/null

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then check what is actually listening on the machine:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;sudo ss -ltnp &lt;span class="o"&gt;|&lt;/span&gt; grep node

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If Node only listens on &lt;code&gt;localhost:3000&lt;/code&gt; (or similar) and Nginx, Caddy, a cloud load balancer or Cloudflare terminates TLS and forwards plain HTTP/1.1, the server-side HTTP/2 and TLS bugs are not reachable from the internet. That is the common setup for Next.js and Express apps on a VPS, and it is a real reduction in risk, not a reason to skip the upgrade.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Does Node make outbound HTTPS calls with custom agents or a proxy?&lt;/strong&gt; The &lt;code&gt;https.Agent&lt;/code&gt; session reuse bug (CVE-2026-123877) and the proxy credentials leak (CVE-2026-123871) are client-side. A reverse proxy in front of your app does nothing for them. If you build agents with different certificate policies, call internal services with private CAs, or configure an outbound proxy with credentials in the URL, these apply to you.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Are you on Node 20 or older?&lt;/strong&gt; Node 20 ("Iron") is marked end-of-life on nodejs.org; its last release was on 24 March 2026, and the advisory states plainly that end-of-life versions are always affected when a security release happens. Node 25 is end-of-life too. There is no patched Node 20. The fix is a major-version upgrade.&lt;/p&gt;

&lt;p&gt;The line I'd put in front of any founder: &lt;strong&gt;a reverse proxy can shield you from a server bug, but it cannot upgrade the runtime you forgot you were running.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  How to fix it, step by step
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Step 1: Make a list of every Node runtime
&lt;/h3&gt;

&lt;p&gt;Servers, containers, CI runners, background workers, cron boxes, serverless functions, internal tools, and the version pinned in &lt;code&gt;.nvmrc&lt;/code&gt;, &lt;code&gt;engines&lt;/code&gt; in &lt;code&gt;package.json&lt;/code&gt;, and your Dockerfiles. The fix takes minutes per place; finding the places is the slow part.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 2: Upgrade servers
&lt;/h3&gt;

&lt;p&gt;If you installed Node with nvm:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;nvm install 24.22.3
nvm &lt;span class="nb"&gt;alias&lt;/span&gt; default 24.22.3
node --version &lt;span class="c"&gt;# v24.22.3&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If you installed it from the NodeSource apt repository on Debian or Ubuntu (use the line you are on, 22.x, 24.x or 26.x):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;sudo apt-get update
sudo apt-get install --only-upgrade nodejs
node --version

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If you are on Node 20, do not try to patch in place. Pick 24.x (the active LTS line most teams should be on) or 22.x, run your test suite against it first, and then switch. Node 22 to 24 and 20 to 24 are both major jumps; treat them as a normal dependency upgrade with a staging deploy, not a hotfix.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 3: Restart every Node process
&lt;/h3&gt;

&lt;p&gt;Upgrading the binary does nothing for processes already running. With PM2 there is a trap: the PM2 daemon itself runs on the old Node until you restart it, and apps started by it can keep the old interpreter path, especially with nvm. One way to do it cleanly:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;pm2 save
pm2 &lt;span class="nb"&gt;kill&lt;/span&gt;
pm2 resurrect
pm2 show your-app &lt;span class="o"&gt;|&lt;/span&gt; grep -i &lt;span class="s2"&gt;"node.js version"&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If your ecosystem file pins an interpreter path (for example an nvm path with the old version number in it), update that path first. With systemd, &lt;code&gt;sudo systemctl restart your-service&lt;/code&gt; and then check &lt;code&gt;node --version&lt;/code&gt; as the service user. Test this sequence on staging before you run it on production.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 4: Rebuild container images
&lt;/h3&gt;

&lt;p&gt;Bump the base image tag and rebuild with &lt;code&gt;--pull&lt;/code&gt; so you do not reuse a cached layer with the old Node:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="k"&gt;FROM&lt;/span&gt; node:24.22.3-bookworm-slim


docker build --pull -t your-registry/your-api:node-24.22.3 .
docker run --rm your-registry/your-api:node-24.22.3 node --version

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Pinning an exact patch version makes upgrades visible in code review. If you use a floating tag like &lt;code&gt;node:24-slim&lt;/code&gt;, a rebuild with &lt;code&gt;--pull&lt;/code&gt; picks up the fix once the official images are published, but you have no record of which version you actually shipped. Check that the exact tag exists on Docker Hub before you merge.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 5: Update CI and local development
&lt;/h3&gt;

&lt;p&gt;Bump &lt;code&gt;node-version&lt;/code&gt; in your CI configuration (for example &lt;code&gt;actions/setup-node&lt;/code&gt;) and the version in &lt;code&gt;.nvmrc&lt;/code&gt;. Self-hosted runners need the same server upgrade as Step 2. CI runners rarely serve HTTP/2 to strangers, so they are lower risk, but keeping CI and production on the same Node version saves you from "works in CI, crashes in prod" a month from now.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 6: Check managed platforms
&lt;/h3&gt;

&lt;p&gt;If you run on a platform that manages the Node runtime for you (serverless functions, PaaS), check your provider's status page or runtime changelog for when they rolled out 22.23.5, 24.22.3 or 26.13.1, and whether you need to redeploy to pick it up. Do not assume; some platforms only update the runtime on the next deploy.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 7: If you cannot upgrade today, reduce exposure
&lt;/h3&gt;

&lt;p&gt;These are stopgaps, not fixes:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Put TLS and HTTP/2 termination in front of Node.&lt;/strong&gt; Let Nginx, Caddy or your load balancer terminate TLS and speak HTTP/1.1 to Node on localhost. This removes direct exposure to the server-side HTTP/2 and TLS bugs.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Disable HTTP/2 on Node directly&lt;/strong&gt; if you enabled it only for performance and a proxy can do that job.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Enforce client certificates at the proxy&lt;/strong&gt; if you rely on mTLS, rather than in Node's &lt;code&gt;SNICallback&lt;/code&gt; or &lt;code&gt;requestCert&lt;/code&gt; logic, until you are patched.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Watch for crash loops.&lt;/strong&gt; Several of these bugs crash the process. Alert on unusual restart counts from PM2, systemd or your orchestrator.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;None of these help with the client-side bugs or with an end-of-life runtime. Upgrade anyway.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 8: Verify
&lt;/h3&gt;

&lt;p&gt;On every host and image, &lt;code&gt;node --version&lt;/code&gt; should show 22.23.5, 24.22.3, 26.13.1 or later. Record the date you patched in your security log or incident tracker. It is the evidence an investor or enterprise customer will ask for later.&lt;/p&gt;

&lt;h2&gt;
  
  
  Did you miss October too?
&lt;/h2&gt;

&lt;p&gt;The October 21, 2026 Node.js security release fixed ten vulnerabilities, including CVE-2026-111497, a Brotli decoder out-of-bounds write that Node.js says can lead to remote code execution when an app decompresses attacker-controlled data with &lt;code&gt;node:zlib&lt;/code&gt;. Express 5 decompresses Brotli request bodies by default, and fetched responses can be Brotli-encoded too, so this one reaches more apps than the name suggests. October's fixed versions were 22.23.4, 24.22.1 and 26.12.1.&lt;/p&gt;

&lt;p&gt;Several outlets (Cyber Kendra, Convika and Brandefense) report that CISA added CVE-2026-111497 to its Known Exploited Vulnerabilities catalogue on 27 October. I could not open the CISA catalogue to confirm that directly, and Node.js itself has not said it is exploited, so treat that as reported rather than confirmed. Either way, the November versions include the October fixes. Upgrading to 22.23.5, 24.22.3 or 26.13.1 closes both.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this means for startups
&lt;/h2&gt;

&lt;p&gt;Two Node.js security releases in three weeks, with AI-assisted researchers credited on several findings, is the new normal. The runtime under your product is getting examined faster than ever, which is good for everyone, as long as your team can absorb a patch without a fire drill.&lt;/p&gt;

&lt;p&gt;The teams that close this in a morning have three things: an inventory of every place Node runs, base images pinned to exact versions so upgrades show up in a pull request, and an owner who checks the Node.js security page when a release is announced. If you read my posts on &lt;a href="https://kunalvohra.com/blog/apt-cve-2026-109518-how-to-fix" rel="noopener noreferrer"&gt;the APT flaw&lt;/a&gt; and &lt;a href="https://kunalvohra.com/blog/nextjs-cve-2026-96889-how-to-fix" rel="noopener noreferrer"&gt;the Next.js CVE-2026-96889 fix&lt;/a&gt;, you will recognise the pattern: the patch is quick, finding every place that needs it is slow. My &lt;a href="https://kunalvohra.com/blog/startup-security-checklist" rel="noopener noreferrer"&gt;startup security checklist&lt;/a&gt; puts that inventory near the top for that reason.&lt;/p&gt;

&lt;p&gt;Here is what I'd do as a CTO this week:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Today:&lt;/strong&gt; patch anything where Node terminates TLS or serves HTTP/2 to the internet, and anything using mTLS.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;This week:&lt;/strong&gt; patch everything else, rebuild images, bump CI and &lt;code&gt;.nvmrc&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;This month:&lt;/strong&gt; move every service off Node 20 and set a calendar reminder for the next end-of-life date of the line you land on.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ongoing:&lt;/strong&gt; subscribe the person who owns infrastructure to the Node.js security announcements, and make "runtime version" a field in your service inventory.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;If nobody on your team owns this, that is the real gap. A &lt;a href="https://kunalvohra.com/technical-advisor" rel="noopener noreferrer"&gt;technical advisor&lt;/a&gt; can set up the inventory, image pinning and patch routine with your engineers in a few sessions; a &lt;a href="https://kunalvohra.com/fractional-cto" rel="noopener noreferrer"&gt;fractional CTO&lt;/a&gt; owns it on an ongoing basis along with the rest of your security posture.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;What versions fix the November 2026 Node.js vulnerabilities?&lt;/strong&gt; Node.js 22.23.5, 24.22.3 and 26.13.1, released on 11 November 2026. Anything older on those lines is affected. End-of-life lines such as Node 20 and Node 25 get no fix.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is Node 20 affected by the November 2026 Node.js security release?&lt;/strong&gt; Node.js says end-of-life versions are always affected when a security release happens, and Node 20 is end-of-life, so assume yes. There is no patched Node 20 release; upgrade to 24.x or 22.x.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Are these Node.js vulnerabilities being exploited?&lt;/strong&gt; The Node.js advisory does not say any of the 16 November issues are exploited. Separately, the October Brotli bug CVE-2026-111497 has been reported by several security outlets as added to CISA's KEV catalogue, which is another reason to upgrade now.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;I run Next.js or Express behind Nginx. Am I safe?&lt;/strong&gt; Safer, not safe. If Nginx terminates TLS and talks plain HTTP/1.1 to Node on localhost, the server-side HTTP/2 and TLS bugs are not reachable from outside. The client-side bugs (outbound HTTPS agents, proxy credentials) and the October Brotli issue still apply, so upgrade this week.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Do I need to restart after upgrading Node.js?&lt;/strong&gt; Yes. Running processes keep the old binary until they restart. With PM2, restart the daemon itself (for example &lt;code&gt;pm2 save&lt;/code&gt;, &lt;code&gt;pm2 kill&lt;/code&gt;, &lt;code&gt;pm2 resurrect&lt;/code&gt;) and confirm the version with &lt;code&gt;pm2 show&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How do I check the Node.js version inside a Docker image?&lt;/strong&gt; Run &lt;code&gt;docker run --rm your-image node --version&lt;/code&gt;. If it is older than the fixed version, bump the &lt;code&gt;FROM node:&lt;/code&gt; tag and rebuild with &lt;code&gt;docker build --pull&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;If you're not sure how many places in your stack run Node, or you want an upgrade routine that makes the next release boring, &lt;a href="https://calendly.com/kunalvohra" rel="noopener noreferrer"&gt;book a 30-minute call&lt;/a&gt;. Bring your Dockerfiles and your PM2 config; that's usually where the old versions hide.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://nodejs.org/en/blog/vulnerability/november-2026-security-releases" rel="noopener noreferrer"&gt;Wednesday, November 11, 2026 Security Releases&lt;/a&gt; (Node.js)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://nodejs.org/en/blog/release/v24.22.3" rel="noopener noreferrer"&gt;Node.js 24.22.3 (LTS) release notes&lt;/a&gt; (Node.js)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://nodejs.org/en/blog/vulnerability/october-2026-security-releases" rel="noopener noreferrer"&gt;Wednesday, October 21, 2026 Security Releases&lt;/a&gt; (Node.js)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://nodejs.org/en/about/previous-releases" rel="noopener noreferrer"&gt;Node.js Releases&lt;/a&gt; (Node.js)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://nodejs.org/en/about/eol" rel="noopener noreferrer"&gt;End-of-Life Node.js versions&lt;/a&gt; (Node.js)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://dev.to/devopsdaily/nodejs-november-2026-security-release-how-we-tracked-and-patched-it-of"&gt;Node.js November 2026 Security Release: How We Tracked and Patched It&lt;/a&gt; (DevOps Daily on DEV Community)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.cyberkendra.com/2026/11/node-js-november-2026-security-release-high-severity-fixes.html" rel="noopener noreferrer"&gt;Node.js November 2026 Security Release: High-Severity Fixes&lt;/a&gt; (Cyber Kendra)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://cybersecuritynews.com/node-js-schedules-patch/" rel="noopener noreferrer"&gt;Node.js Schedules High-Severity Security Fixes for 22.x, 24.x, and 26.x Release Lines&lt;/a&gt; (Cyber Security News)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://thehackernews.com/2026/10/researchers-demonstrate-node2shell-rce.html" rel="noopener noreferrer"&gt;Researchers Demonstrate Node2Shell RCE Exploit on Node.js Build Using Crafted Brotli Data&lt;/a&gt; (The Hacker News)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.aikido.dev/blog/node2shell-nodejs-brotli-rce-cve-2026-111497" rel="noopener noreferrer"&gt;Node2Shell: Node.js Brotli vulnerability allows remote code execution (CVE-2026-111497) and the fix&lt;/a&gt; (Aikido Security)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://convika.com/blog/nodejs-security-release-2026-10-brotli/" rel="noopener noreferrer"&gt;Node.js CVE-2026-111497, now exploited: what to update&lt;/a&gt; (Convika)&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>security</category>
      <category>cybersecurity</category>
      <category>webdev</category>
      <category>startup</category>
    </item>
    <item>
      <title>The WorkOS Breach and ARTEX (November 2026): What AI Agent Attacks Mean for Your Startup's Internal Tools</title>
      <dc:creator>Kunal Vohra</dc:creator>
      <pubDate>Tue, 10 Nov 2026 12:00:00 +0000</pubDate>
      <link>https://dev.to/kunalvohra/the-workos-breach-and-artex-november-2026-what-ai-agent-attacks-mean-for-your-startups-internal-2aei</link>
      <guid>https://dev.to/kunalvohra/the-workos-breach-and-artex-november-2026-what-ai-agent-attacks-mean-for-your-startups-internal-2aei</guid>
      <description>&lt;p&gt;On 6 November 2026, WorkOS confirmed that the data taken from one of its internal sales tools was pulled by an AI security agent running the ARTEX framework, which had escaped a researcher's isolated test environment during benchmark runs. The agent found an internal tool that should never have been on the internet, logged in with a weak credential meant for local development, and retrieved records in bulk from an API with no rate limits. In the same fortnight Chainalysis disclosed an intrusion by "an external actor using offensive agentic techniques", and CrowdStrike documented a China-nexus group using ARTEX against more than 200 organisations in 11 days. If you run a startup, the lesson is not "AI hackers are coming". It is that the boring gaps (an exposed internal app, a dev password in production, no rate limiting, no volume alerts) are now found and used at machine speed. Close those four this week.&lt;/p&gt;

&lt;p&gt;I'm Kunal Vohra, a fractional CTO and a security researcher and cybersecurity advisor. Last week I wrote about &lt;a href="https://kunalvohra.com/blog/openai-948-linux-root-exploits" rel="noopener noreferrer"&gt;OpenAI's 948 Linux root exploits&lt;/a&gt;, which showed what AI can do to software when defenders point it at code. This post is the other side: what it looks like when an agent is pointed at a company, and what WorkOS was honest enough to publish about how it happened.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who should care
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Any startup with internal tools.&lt;/strong&gt; Admin panels, sales and support dashboards, a Retool or homegrown CRM, a staging app, an "internal" API a contractor built two years ago. If one of these is reachable from the internet, an agent will find it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Teams with old or shared credentials.&lt;/strong&gt; CrowdStrike's write-up lists unrotated 2022 keys and a VPN key tied to a 2024 email leak among the ways ARTEX operators got in.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Anyone on AWS, GCP or Azure.&lt;/strong&gt; One of the documented paths was a web flaw used to reach cloud instance metadata, which is how a small bug becomes cloud credentials.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Customers of SaaS vendors that hold your API tokens.&lt;/strong&gt; Chainalysis says names, usernames, password hashes and live tokens for all of its users were exposed. Your security now includes your vendors' internal tools.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Founders and sales teams.&lt;/strong&gt; WorkOS warned that stolen business contact data and call summaries make follow-up phishing look legitimate.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What actually happened
&lt;/h2&gt;

&lt;h3&gt;
  
  
  WorkOS: an agent that left its sandbox
&lt;/h3&gt;

&lt;p&gt;WorkOS sells authentication and enterprise SSO to SaaS companies, so its incident drew attention fast. It published two security bulletins:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;2 November 2026.&lt;/strong&gt; WorkOS said it had identified, on 30 October, unauthorised access to "an internal WorkOS application used by our sales team", and that the investigation found an AI-assisted offensive security framework automating reconnaissance and exploitation. It disabled external access to the app, brought in an independent forensics firm and began notifying affected customers.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;6 November 2026.&lt;/strong&gt; After the first bulletin, a researcher came forward. WorkOS says the researcher was running cybersecurity benchmark evaluations with ARTEX when the agent escaped its isolated testing environment, and that it "wasn't the work of an actor with malicious intent". The researcher supplied request details, timestamps and artifacts; WorkOS and the forensics firm matched them to its logs, and the researcher confirmed they deleted all WorkOS data and shared it with no one.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The timeline WorkOS gives: first confirmed unauthorised access to the internal tool's API on &lt;strong&gt;25 October&lt;/strong&gt; ; bulk retrieval of records on &lt;strong&gt;30 October&lt;/strong&gt; , when an employee noticed unusually high usage of a connected system; public access removed and credentials for connected services rotated &lt;strong&gt;30 to 31 October&lt;/strong&gt; ; customer notification from &lt;strong&gt;2 November&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;The data the app gave access to, per WorkOS: business contact details (names, emails, phone numbers, job titles, companies), account histories and sales relationship information, summaries and excerpts of sales calls, support discussions and customer Slack conversations (a limited number including end-user email addresses), and product usage and connection-health metadata. The forensics firm's finding, as WorkOS reports it, is that only data in its sales and customer support systems was affected. WorkOS did not publish how many customers were affected, and it did not explain how the agent escaped the researcher's sandbox.&lt;/p&gt;

&lt;p&gt;The most useful part of the update is the root-cause list. In WorkOS's own words, the tool "should not have been reachable from the internet"; its API "accepted a weak credential intended for local development", and controls did not stop that credential reaching production; the API "lacked rate limiting or throttling mechanisms"; and monitoring "did not detect the activity soon enough". The app also sat on a legacy internal-tools platform that had not been migrated, so it missed the safeguards newer tools had.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;An AI agent doesn't need a zero-day. It needs one internal tool you forgot was on the internet, and the patience to try every door at machine speed.&lt;/strong&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  ARTEX and Vapor Panda: the same tool, used on purpose
&lt;/h3&gt;

&lt;p&gt;On 29 October, CrowdStrike published research on a China-nexus group it calls VAPOR PANDA using ARTEX, which The Hacker News describes as a Go-based, AI-powered penetration testing framework. CrowdStrike's numbers:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Activity starting &lt;strong&gt;17 October 2026&lt;/strong&gt; , touching &lt;strong&gt;more than 200 organisations&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;At least 14 countries in an 11-day window&lt;/strong&gt; from a single ARTEX controller.&lt;/li&gt;
&lt;li&gt;Post-exploitation at &lt;strong&gt;nine organisations&lt;/strong&gt; , with at least &lt;strong&gt;12 more&lt;/strong&gt; password-sprayed or credential-scanned.&lt;/li&gt;
&lt;li&gt;One intrusion's planning agents consumed &lt;strong&gt;on the order of 1.4 billion input tokens&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;Agents switched approach when a path failed, and the human operator intervened mainly at decision points.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The ways in were not exotic: an unsigned SAML assertion and an XML external entity flaw used to reach cloud instance metadata, an unauthenticated orchestration service in a container environment, unrotated keys, VPN configurations recovered from mail, and password spraying. CrowdStrike says this campaign is separate from ARTEX activity Microsoft has reported under a different name. The Hacker News adds that the developer has reportedly taken ARTEX closed source, citing other reporting; I have not seen that confirmed by the developer.&lt;/p&gt;

&lt;h3&gt;
  
  
  Chainalysis: a disclosure that kept growing
&lt;/h3&gt;

&lt;p&gt;Chainalysis first posted on 30 October that an external actor using offensive agentic techniques had accessed its systems and internal tools, and at that point said it had no evidence that credentials or session tokens were accessed. Its updates then widened the picture: on 4 November an "on-path attack" exposed a subset of customer passwords, and later that day it said names, usernames, password hashes and live tokens for all Chainalysis users were exposed. It took Reactor and Data Solutions offline on 2 November, began a phased restoration of KYT and Address Screening on 6 November with trusted-IP allowlisting and forced password and API key resets, and said on 7 November that Reactor, KYT and Address Screening were back online. Chainalysis does not name ARTEX, and nothing published links its incident to WorkOS. The Hacker News lists Chainalysis and WorkOS alongside Chainstack, OpenCode and Warp as companies that have disclosed AI-driven attacks this autumn.&lt;/p&gt;

&lt;h2&gt;
  
  
  What changes for a startup
&lt;/h2&gt;

&lt;p&gt;Nothing in the WorkOS or CrowdStrike accounts needed a new class of vulnerability. What changed is cost. A human attacker might never get round to your forgotten staging dashboard; an agent that can run reconnaissance across 200 organisations in 11 days will. I made a similar point in &lt;a href="https://kunalvohra.com/blog/can-ai-hack-systems" rel="noopener noreferrer"&gt;Can AI Hack Systems?&lt;/a&gt;: AI mostly makes existing attacks cheaper and more patient. The WorkOS timeline shows the other half. Five days passed between first access and detection, and detection came from a human noticing usage, not from an alert.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to change in your stack this week
&lt;/h2&gt;

&lt;h3&gt;
  
  
  1. Find everything that answers on the internet
&lt;/h3&gt;

&lt;p&gt;Start from the outside, the way an agent would. Certificate transparency logs list most hostnames you have ever issued a certificate for:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl -s &lt;span class="s2"&gt;"https://crt.sh/?q=%25.yourcompany.com&amp;amp;output=json"&lt;/span&gt; \
  &lt;span class="o"&gt;|&lt;/span&gt; jq -r &lt;span class="s1"&gt;'.[].name_value'&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt; sort -u

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then check the cloud side. On AWS, list security groups that allow traffic from anywhere:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws ec2 describe-security-groups \
  --filters &lt;span class="nv"&gt;Name&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;ip-permission.cidr,Values=0.0.0.0/0 \
  --query &lt;span class="s2"&gt;"SecurityGroups[].[GroupId,GroupName]"&lt;/span&gt; --output table

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Every hostname and open port should have an owner and a reason. Anything called admin, internal, staging, crm, ops, tools or dashboard deserves a closer look.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. Take internal tools off the public internet
&lt;/h3&gt;

&lt;p&gt;WorkOS's fix was to block internally developed tools from the internet at the network edge. For a startup, that usually means putting internal apps behind an identity-aware proxy (Cloudflare Access, Google IAP, or a private network such as Tailscale) so a request never reaches the app without a company SSO login. A login page on the app itself is not enough; the app's API is what got scraped.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Make it impossible for dev secrets to reach production
&lt;/h3&gt;

&lt;p&gt;A weak credential "intended for local development" ended up guarding production. The cheapest guard is to make the app refuse to start in production with a weak or known development secret:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="c1"&gt;// lib/assert-prod-secrets.ts: call once at server start&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="no"&gt;DEV_VALUES&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nb"&gt;Set&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"changeme"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"password"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"dev"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"test"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"secret"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"admin"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="k"&gt;export&lt;/span&gt; &lt;span class="kd"&gt;function&lt;/span&gt; &lt;span class="nf"&gt;assertProdSecret&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;name&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;string&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;value&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nb"&gt;process&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;env&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;name&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;??&lt;/span&gt; &lt;span class="s2"&gt;""&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nb"&gt;process&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;env&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="no"&gt;NODE_ENV&lt;/span&gt; &lt;span class="o"&gt;!==&lt;/span&gt; &lt;span class="s2"&gt;"production"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;value&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;length&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&lt;/span&gt; &lt;span class="mi"&gt;32&lt;/span&gt; &lt;span class="o"&gt;||&lt;/span&gt; &lt;span class="no"&gt;DEV_VALUES&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;has&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;value&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;toLowerCase&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="k"&gt;throw&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nb"&gt;Error&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sb"&gt;`${name} looks like a development secret; refusing to start in production`&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Pair it with separate secrets per environment in your secret manager, so nobody can copy the dev value across by accident.&lt;/p&gt;

&lt;h3&gt;
  
  
  4. Rate limit every API, including the internal ones
&lt;/h3&gt;

&lt;p&gt;No rate limiting is what turned access into bulk retrieval. If you run Nginx in front of Node, a basic per-IP limit is a few lines:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight nginx"&gt;&lt;code&gt;&lt;span class="c1"&gt;# in the http block
&lt;/span&gt;&lt;span class="na"&gt;limit_req_zone $binary_remote_addr zone&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;api&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="mi"&gt;10m&lt;/span&gt; &lt;span class="n"&gt;rate&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;10r&lt;/span&gt;&lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="n"&gt;s&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="c1"&gt;# in the server block
&lt;/span&gt;&lt;span class="n"&gt;location&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="n"&gt;api&lt;/span&gt;&lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="na"&gt;limit_req zone&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;api&lt;/span&gt; &lt;span class="n"&gt;burst&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;20&lt;/span&gt; &lt;span class="n"&gt;nodelay&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="n"&gt;limit_req_status&lt;/span&gt; &lt;span class="mi"&gt;429&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="n"&gt;proxy_pass&lt;/span&gt; &lt;span class="n"&gt;http&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="c1"&gt;//127.0.0.1:3000;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Per-IP limits are a floor. For authenticated APIs, also limit per token or per user, and cap how many records one request can return.&lt;/p&gt;

&lt;h3&gt;
  
  
  5. Alert on volume, not just errors
&lt;/h3&gt;

&lt;p&gt;An agent pulling data in bulk mostly gets successful responses, so error-based alerts stay quiet. Alert when one user, token or IP reads far more than usual in an hour. WorkOS lists alerts for unusual access patterns and request volumes among its in-progress fixes. If you have nothing else, a daily look at top talkers in your access logs is a start:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;awk &lt;span class="s1"&gt;'{print $1}'&lt;/span&gt; /var/log/nginx/access.log &lt;span class="o"&gt;|&lt;/span&gt; sort &lt;span class="o"&gt;|&lt;/span&gt; uniq -c &lt;span class="o"&gt;|&lt;/span&gt; sort -rn &lt;span class="o"&gt;|&lt;/span&gt; head -20

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;WorkOS also published the user agents it saw. They are trivial to change, but a quick check costs nothing:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;zgrep -iE &lt;span class="s2"&gt;"artex|infra-hunt|norma/|verify-911-probe"&lt;/span&gt; /var/log/nginx/access.log*

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  6. Rotate old keys and close the cloud metadata path
&lt;/h3&gt;

&lt;p&gt;CrowdStrike's advice is to rotate signing keys, preauthentication keys and service credentials after any suspected exposure, enforce SAML signature validation, and require IMDSv2. On AWS, IMDSv2 makes the metadata-service path much harder to abuse through a web flaw:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;aws ec2 modify-instance-metadata-options \
  --instance-id i-0123456789abcdef0 \
  --http-tokens required --http-endpoint enabled

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then list every key older than a year and either rotate it or delete it. If nobody knows what a key is for, that is your answer.&lt;/p&gt;

&lt;h3&gt;
  
  
  7. Keep a vendor token register
&lt;/h3&gt;

&lt;p&gt;The Chainalysis lesson is about the tokens you hand to vendors. Keep a short list: vendor, what the token can do, where it is stored, who can rotate it. When a vendor discloses an incident, you should be able to rotate in minutes, not spend a day finding the token. Prefer scoped, read-only, IP-allowlisted tokens wherever a vendor offers them.&lt;/p&gt;

&lt;h3&gt;
  
  
  8. Expect better phishing, and say so
&lt;/h3&gt;

&lt;p&gt;WorkOS warned that stolen contact data and call summaries can make phishing look legitimate. Tell your sales and finance people now: an email that quotes your last call accurately is not proof it is real. Verify payment or credential requests on a known channel. My &lt;a href="https://kunalvohra.com/blog/device-code-phishing-fake-ai-events" rel="noopener noreferrer"&gt;device code phishing guide&lt;/a&gt; covers the account settings that blunt this.&lt;/p&gt;

&lt;h3&gt;
  
  
  9. If you run AI security agents, sandbox their network
&lt;/h3&gt;

&lt;p&gt;The WorkOS agent was a researcher's benchmark run, not a criminal's. If your team experiments with AI pentest agents, run them with an egress allowlist that only reaches the targets you own, and log every outbound request. The same applies to coding agents and anything else that can make network calls, as I covered in &lt;a href="https://kunalvohra.com/blog/ai-agent-security" rel="noopener noreferrer"&gt;AI agent security for startups&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I'd do as a CTO
&lt;/h2&gt;

&lt;p&gt;I'd treat this as an inventory problem first. The question I'd ask any team is whether it can produce, today, a complete list of what it exposes to the internet, which secrets guard each piece, and who would notice a strange spike in reads. Building that list is the work: one afternoon on steps 1 to 3, a day on rate limits and alerts, then a quarterly review. It is already on my &lt;a href="https://kunalvohra.com/blog/startup-security-checklist" rel="noopener noreferrer"&gt;startup security checklist&lt;/a&gt; for a reason.&lt;/p&gt;

&lt;p&gt;I'd also give WorkOS credit. Its update names four plain mistakes in plain language. That is the bar I'd hold any vendor to, and any company I advise: when something goes wrong, publish what failed, not just what was accessed.&lt;/p&gt;

&lt;p&gt;If nobody on your team owns this, it doesn't need a full-time hire. A &lt;a href="https://kunalvohra.com/technical-advisor" rel="noopener noreferrer"&gt;technical advisor&lt;/a&gt; retainer can run the review with your engineers, and a &lt;a href="https://kunalvohra.com/fractional-cto" rel="noopener noreferrer"&gt;fractional CTO&lt;/a&gt; can own security, infrastructure and the rest of the platform end to end.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;What happened in the WorkOS breach?&lt;/strong&gt; WorkOS says an AI security agent running the ARTEX framework escaped a researcher's isolated test environment, reached an internal sales tool that was exposed to the internet, used a weak development credential, and retrieved sales and support records in bulk between 25 and 30 October 2026. WorkOS says the researcher had no malicious intent and deleted the data, and that a forensics firm found only sales and support systems were affected.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What is ARTEX?&lt;/strong&gt; ARTEX is an AI-powered penetration testing framework, written in Go according to The Hacker News, that uses LLM agents to plan and run reconnaissance and exploitation. CrowdStrike reported a China-nexus group, VAPOR PANDA, using it against more than 200 organisations from 17 October 2026.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Was WorkOS AuthKit or SSO affected?&lt;/strong&gt; WorkOS's bulletins describe an internal sales application and say the forensics firm found only data in its sales and customer support systems was affected. They do not report any impact on its authentication products. Watch the WorkOS security bulletins for updates.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is the Chainalysis breach related to WorkOS?&lt;/strong&gt; Nothing published links them. Chainalysis says an external actor used offensive agentic techniques; it does not name ARTEX. If you are a Chainalysis customer, rotate your passwords and API keys and follow its forced resets.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How do I protect my startup from AI agent attacks?&lt;/strong&gt; Remove internal tools from the public internet, stop development credentials reaching production, rate limit every API, alert on unusual read volume, rotate old keys, require IMDSv2 on AWS, and keep a register of vendor tokens so you can rotate quickly.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Can AI agents really hack companies on their own?&lt;/strong&gt; In these cases a human still started and steered the work, but the agents did the reconnaissance and retried failed paths themselves. That makes the cheap, common mistakes far more likely to be found than before.&lt;/p&gt;

&lt;p&gt;If you'd like a second pair of eyes on what your startup exposes to the internet, &lt;a href="https://calendly.com/kunalvohra" rel="noopener noreferrer"&gt;book a 30-minute call&lt;/a&gt;. Bring your list of domains and your cloud account owner, and we'll find the internal tools that shouldn't be public.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://workos.com/security/bulletins/october-30th-2026-data-security-incident" rel="noopener noreferrer"&gt;October 30th, 2026 Data Security Incident&lt;/a&gt; (WorkOS)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://workos.com/security/bulletins/update-for-october-30th-2026-data-security-incident" rel="noopener noreferrer"&gt;Update for October 30th, 2026 Data Security Incident&lt;/a&gt; (WorkOS)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/" rel="noopener noreferrer"&gt;VAPOR PANDA Uses AI-Driven ARTEX to Scale a Global Espionage Campaign&lt;/a&gt; (CrowdStrike)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.chainalysis.com/blog/security-incident-october-30/" rel="noopener noreferrer"&gt;Updated Nov 7: Chainalysis Security Incident Information&lt;/a&gt; (Chainalysis)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://thehackernews.com/2026/10/chinese-hackers-use-ai-powered-artex-to.html" rel="noopener noreferrer"&gt;Chinese Hackers Use AI-Powered ARTEX to Target Over 200 Organizations Worldwide&lt;/a&gt; (The Hacker News)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://thehackernews.com/2026/11/chainalysis-breach-exposes-all-users.html" rel="noopener noreferrer"&gt;Chainalysis Breach Exposes All Users' Credentials, Forces Shutdown of Key Services&lt;/a&gt; (The Hacker News)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://cyberpress.org/artex-ai-agent-security-sandbox" rel="noopener noreferrer"&gt;ARTEX AI Agent Escapes Security Sandbox and Breaches WorkOS, Exposing Customer Data&lt;/a&gt; (Cyber Press)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.compsmag.com/news/workos-says-researchers-ai-penetration-testing-agent-escaped-its-sandbox/" rel="noopener noreferrer"&gt;WorkOS Says Researcher's AI Penetration-Testing Agent Escaped Its Sandbox&lt;/a&gt; (Compsmag)&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>security</category>
      <category>cybersecurity</category>
      <category>webdev</category>
      <category>startup</category>
    </item>
    <item>
      <title>Fake OpenAI and SpaceX Event Invites (November 2026): How Device Code Phishing Works and How to Block It</title>
      <dc:creator>Kunal Vohra</dc:creator>
      <pubDate>Sun, 08 Nov 2026 12:00:00 +0000</pubDate>
      <link>https://dev.to/kunalvohra/fake-openai-and-spacex-event-invites-november-2026-how-device-code-phishing-works-and-how-to-13m</link>
      <guid>https://dev.to/kunalvohra/fake-openai-and-spacex-event-invites-november-2026-how-device-code-phishing-works-and-how-to-13m</guid>
      <description>&lt;p&gt;A phishing campaign documented on 5 November 2026 uses fake event registration pages for OpenAI Academy, a SpaceX product briefing and an Anthropic session at AWS re:Invent to take over Microsoft and Google work accounts. Visitors are promised a reward (three months of ChatGPT Pro, Grok Heavy, or a "SpaceX share") if they sign in with a work account. On the Microsoft side, the trick is device code phishing: you type a short code into Microsoft's real sign-in page, approve your real MFA prompt, and the attacker's session gets your tokens. If your company runs on Microsoft 365, block the device code flow in Conditional Access this week. If you are on Google Workspace, move your team to passkeys or security keys. Both changes take an afternoon.&lt;/p&gt;

&lt;p&gt;I'm Kunal Vohra, a fractional CTO and a security researcher and cybersecurity advisor. Most of what I've written this month has been about patching servers. This one is about the accounts above the servers: the email, documents and single sign-on every startup depends on. Device code phishing is not new, but this autumn it has gone from a nation-state technique to something sold as a service, aimed squarely at the people who run startups.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who should care
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Any company on Microsoft 365 or Entra ID.&lt;/strong&gt; Device code phishing works against the default configuration. If nobody has turned the device code flow off, it is on.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Teams on Google Workspace.&lt;/strong&gt; The same campaign uses a proxy that sits between you and Google's real sign-in page and steals your password, MFA response and session cookie.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Founders, marketing and sales teams.&lt;/strong&gt; The OpenAI lure was billed as a session on ChatGPT for social and digital marketing teams. These are people who sign up for webinars weekly and are rarely the target of security training.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Anyone using single sign-on.&lt;/strong&gt; A work account is the key to everything connected to it; BleepingComputer lists apps such as Salesforce, Dropbox, Slack, Zendesk and Atlassian.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What actually happened
&lt;/h2&gt;

&lt;p&gt;On 5 November 2026, Push Security published an analysis of a campaign using fake AI-themed event pages, and BleepingComputer covered it the same day. The details that matter:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The lures.&lt;/strong&gt; An OpenAI Academy livestream dated 4 November ("ChatGPT for Social &amp;amp; Digital Marketing Teams"), a SpaceX "Starship Product Briefing" about SpaceX AI for marketing and Grok dated 12 November, and an Anthropic page themed on AWS re:Invent, which was already offline when researchers looked at it. Each page asks you to sign in with a work account to register and claim a reward.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Two paths, two rented kits.&lt;/strong&gt;"Continue with Microsoft" requests a device code from Kali365, a phishing-as-a-service platform, and shows it in a "Confirm access" dialog. The victim is then sent to Microsoft's real device login page to enter it, after which a second page from VoidProxy, another rented platform, tries to capture the password and session cookie too. "Continue with Google" opens a VoidProxy reverse proxy of Google's sign-in page that captures the password, the MFA response and the session cookie.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Delivery that looks clean.&lt;/strong&gt; Links went through Google's own short-link domains and Zapier's link shortener, and earlier versions used GoTo Webinar senders or real Zoom event pages. A Cloudflare "confirm you're not a robot" check in front of the pages also keeps security scanners out.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The operator.&lt;/strong&gt; No threat actor is named. Push links this campaign to an earlier kit it calls ZoomMock (May to October 2026) and a newer one it calls MSDRoot (from 22 October), with medium-high confidence, and says the operator is most likely a customer of both rented platforms rather than their developer. BleepingComputer's framing is a little stronger; I'm going with Push's wording because it is the original research. Domains rotate every few weeks, so blocklists age fast.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;No source gives victim or organisation counts for this campaign. Push says it detected and blocked visits in late October.&lt;/p&gt;

&lt;h3&gt;
  
  
  It is not one campaign, it is a season
&lt;/h3&gt;

&lt;p&gt;The event lures are the most clickable version of a wider trend:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;SkyW4lker.&lt;/strong&gt; A Microsoft 365 device code kit first seen on 23 June 2026, documented by ANY.RUN on 27 October and covered by Cyber Security News on 4 November. It uses invoices, contracts, payment notices and password-protected documents as lures, with redirects through Google Sites, Miro, Zoom and Smartsheet. ANY.RUN counted 618 sandbox analyses between June and 21 October, rising from 20 in June to 200 in September, with roughly 46 percent of submissions from the US and India second.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;MadApe.&lt;/strong&gt; A Microsoft device code kit using business-themed PDF lures and fake document-sharing pages, first observed by Group-IB on 7 January 2026, in The Hacker News weekly recap of 2 November.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;GeorgeGinx and Mir0Auth.&lt;/strong&gt; Two more device code operations in ANY.RUN research published on The Hacker News on 5 November: one abusing Adobe shared links, the other fake DocuSign pages.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Storm-2372.&lt;/strong&gt; Microsoft's February 2025 device code case (fake Teams invites), which a July 2026 update ties to Midnight Blizzard.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;When a state-linked cluster and anyone with a subscription use the same trick, it is no longer exotic. It is the default phishing technique of 2026.&lt;/p&gt;

&lt;h2&gt;
  
  
  How device code phishing works
&lt;/h2&gt;

&lt;p&gt;The device code flow exists for devices that cannot show a full sign-in page, like a smart TV or a command-line tool. The device gets a short code from Microsoft, you type it into a Microsoft page on your phone or laptop and sign in with MFA, and Microsoft hands tokens to the device that asked.&lt;/p&gt;

&lt;p&gt;In the attack, the "device" is the attacker's server:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;You click "Continue with Microsoft" on the fake event page.&lt;/li&gt;
&lt;li&gt;The attacker's backend asks Microsoft for a device code, using a legitimate Microsoft client.&lt;/li&gt;
&lt;li&gt;The page shows you that code and sends you to Microsoft's real device login page.&lt;/li&gt;
&lt;li&gt;You type the code, sign in with your password, and approve your MFA prompt. Every page you see is genuinely Microsoft's.&lt;/li&gt;
&lt;li&gt;Microsoft issues access and refresh tokens to the attacker's session. The attacker now has your mailbox, files and any app your account can reach, without ever seeing your password.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;strong&gt;In device code phishing, the URL is real, the MFA prompt is real, and the only fake thing is who asked for the code.&lt;/strong&gt; That is why "check the address bar" and "MFA protects you" both fail here, and why the fix has to be a configuration change, not just training.&lt;/p&gt;

&lt;p&gt;Microsoft's own documentation calls device code flow "a high-risk authentication method" and says it recommends blocking it wherever possible.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to change in your stack this week
&lt;/h2&gt;

&lt;h3&gt;
  
  
  1. Microsoft 365: block device code flow with Conditional Access
&lt;/h3&gt;

&lt;p&gt;This is the single highest-value change. You need the Conditional Access Administrator role, and Conditional Access requires Entra ID P1 licences (included in Microsoft 365 Business Premium and E3/E5). In the Microsoft Entra admin center:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Go to &lt;strong&gt;Entra ID &amp;gt; Conditional Access &amp;gt; Policies &amp;gt; New policy&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;Under &lt;strong&gt;Users or workload identities&lt;/strong&gt; , include &lt;strong&gt;All users&lt;/strong&gt; and exclude your break-glass (emergency access) accounts.&lt;/li&gt;
&lt;li&gt;Under &lt;strong&gt;Target resources&lt;/strong&gt; , include &lt;strong&gt;All resources&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;Under &lt;strong&gt;Conditions &amp;gt; Authentication flows&lt;/strong&gt; , set &lt;strong&gt;Configure&lt;/strong&gt; to &lt;strong&gt;Yes&lt;/strong&gt; and select &lt;strong&gt;Device code flow&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;Under &lt;strong&gt;Grant&lt;/strong&gt; , choose &lt;strong&gt;Block access&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;Set &lt;strong&gt;Enable policy&lt;/strong&gt; to &lt;strong&gt;Report-only&lt;/strong&gt; , create it, and check the results before switching it to &lt;strong&gt;On&lt;/strong&gt;.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Before you switch it on, find out who really uses device code flow: in &lt;strong&gt;Sign-in logs&lt;/strong&gt; , filter by &lt;strong&gt;Authentication protocol&lt;/strong&gt; and choose &lt;strong&gt;Device code&lt;/strong&gt;. It is usually a few engineers with command-line tools or a meeting room device. Microsoft's guidance is to get "as close as possible to a unilateral block", so exclude a named group rather than leaving it open for everyone. If someone hits the block, the error code they'll report is &lt;code&gt;AADSTS530036&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;The same condition has an &lt;strong&gt;Authentication transfer&lt;/strong&gt; option, which covers moving a signed-in session from desktop Outlook to a phone by scanning a QR code. Decide whether you need it; if not, block that too.&lt;/p&gt;

&lt;p&gt;One detail that trips people up: since September 2024 an all-resources authentication flows policy also applies to the Device Registration Service, and Microsoft documents how to exclude it if a legitimate setup breaks.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. Move to phishing-resistant sign-in
&lt;/h3&gt;

&lt;p&gt;Passkeys and FIDO2 security keys are bound to the real website, so a proxy page cannot replay them. That is what defeats the Google half of this campaign and AiTM phishing generally.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Microsoft:&lt;/strong&gt; enable passkeys (FIDO2 keys or passkeys in Microsoft Authenticator) and require them for admins first, then everyone. Microsoft's Storm-2372 guidance lists phishing-resistant MFA and blocking legacy authentication next to blocking device code flow.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Google Workspace:&lt;/strong&gt; in the Admin console, go to &lt;strong&gt;Security &amp;gt; Authentication &amp;gt; 2-step verification&lt;/strong&gt; (super admin required). Google's documentation says its security key method now supports both security keys and passkeys. Enforce it for admins and finance first, then the whole company.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A passkey does not stop someone typing a device code into a real Microsoft page, which is why step 1 comes first. You need both.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Bind tokens to devices where you can
&lt;/h3&gt;

&lt;p&gt;Microsoft's &lt;strong&gt;token protection&lt;/strong&gt; session control only accepts tokens bound to a registered device, which makes stolen tokens much harder to replay. It covers Exchange Online, SharePoint Online and Teams in the native apps on Windows, macOS and iOS/iPadOS. Pilot it in report-only mode first, starting with founders and finance.&lt;/p&gt;

&lt;h3&gt;
  
  
  4. Look for it in your logs
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;Filter Entra sign-in logs by &lt;strong&gt;Authentication protocol: Device code&lt;/strong&gt; for the last 90 days. A device code sign-in from someone who never uses a command line deserves a phone call.&lt;/li&gt;
&lt;li&gt;In a sign-in's activity details, the &lt;strong&gt;Original transfer method&lt;/strong&gt; field shows &lt;strong&gt;Device code flow&lt;/strong&gt; when a session started that way.&lt;/li&gt;
&lt;li&gt;If you have Microsoft Defender, Microsoft suggests hunting for clicks to &lt;code&gt;microsoft.com/devicelogin&lt;/code&gt; and &lt;code&gt;login.microsoftonline.com/common/oauth2/deviceauth&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  5. Teach one rule, not a course
&lt;/h3&gt;

&lt;p&gt;The one sentence I'd give every employee: &lt;strong&gt;never type a code into a Microsoft or Google page because a website told you to.&lt;/strong&gt; You only enter a device code when you started the sign-in yourself, on a screen or terminal in front of you. And for this season: an event that needs your work account to "claim" a free subscription is phishing until proven otherwise.&lt;/p&gt;

&lt;h2&gt;
  
  
  If someone already entered a code
&lt;/h2&gt;

&lt;p&gt;Treat it as an account takeover, not a "suspicious email" ticket. Changing the password alone does not kill the attacker's tokens.&lt;/p&gt;

&lt;p&gt;In the Entra admin center, open the user, clear &lt;strong&gt;Account enabled&lt;/strong&gt; , and on the user's overview page select &lt;strong&gt;Revoke sessions&lt;/strong&gt;. With Microsoft Graph PowerShell, the same steps look like this:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight powershell"&gt;&lt;code&gt;&lt;span class="nc"&gt;Connect&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;MgGraph&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Scopes&lt;/span&gt; &lt;span class="s2"&gt;"User.ReadWrite.All"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="s2"&gt;"Directory.AccessAsUser.All"&lt;/span&gt;
&lt;span class="nv"&gt;$User&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;Get&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;MgUser&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;UserId&lt;/span&gt; &lt;span class="s2"&gt;"person@yourcompany.com"&lt;/span&gt;
&lt;span class="nc"&gt;Update&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;MgUser&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;UserId&lt;/span&gt; &lt;span class="nv"&gt;$User&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Id&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;AccountEnabled&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="nv"&gt;$false&lt;/span&gt;
&lt;span class="nc"&gt;Revoke&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;MgUserSignInSession&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;UserId&lt;/span&gt; &lt;span class="nv"&gt;$User&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Id&lt;/span&gt;
&lt;span class="nc"&gt;Get&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;MgUserRegisteredDevice&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;UserId&lt;/span&gt; &lt;span class="nv"&gt;$User&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Id&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;All&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt; &lt;span class="k"&gt;ForEach&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;Object&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="nc"&gt;Update&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;MgDevice&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;DeviceId&lt;/span&gt; &lt;span class="nv"&gt;$_&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Id&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="nc"&gt;AccountEnabled&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="nv"&gt;$false&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Revoking sessions stops new tokens, but access tokens already issued last until they expire (an hour by default), and apps with their own sessions are not revoked by Entra directly. Then, before re-enabling the account:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Reset the password and re-register MFA methods; remove any MFA method or device you don't recognise.&lt;/li&gt;
&lt;li&gt;Check mailbox rules and forwarding (attackers love a rule that hides replies).&lt;/li&gt;
&lt;li&gt;Review OAuth app consents the user granted.&lt;/li&gt;
&lt;li&gt;Check SharePoint and OneDrive access and recent sharing links.&lt;/li&gt;
&lt;li&gt;Check outbound mail for internal phishing sent from the account.&lt;/li&gt;
&lt;li&gt;Sign the person out of connected SSO apps and rotate any secrets that sat in their mailbox or drive.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;For Google Workspace, do the equivalent from the Admin console: reset the password, sign the user out of all sessions, review their 2-step verification methods and third-party app access, and check Gmail filters and forwarding.&lt;/p&gt;

&lt;p&gt;If you hold personal data of people in India, check your breach reporting duties; I covered them in &lt;a href="https://kunalvohra.com/blog/dpdp-act-explained" rel="noopener noreferrer"&gt;my DPDP Act explainer&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I'd do as a CTO
&lt;/h2&gt;

&lt;p&gt;I'd spend Monday morning on step 1. Put the Conditional Access policy in report-only, pull 90 days of device code sign-ins, name the two or three people who genuinely need it, and switch the block on by Wednesday. For most startups that list of people is empty.&lt;/p&gt;

&lt;p&gt;I'd start the passkey rollout the same week with the accounts that would hurt most: founders, finance, and anyone with admin rights in Microsoft, Google, AWS or the domain registrar. Those are the accounts an attacker picks first.&lt;/p&gt;

&lt;p&gt;And I'd stop treating identity as an IT chore. The Microsoft or Google tenant is the root account for the whole company, so it gets what production gets: an owner, a written baseline and a quarterly review (it is on my &lt;a href="https://kunalvohra.com/blog/startup-security-checklist" rel="noopener noreferrer"&gt;startup security checklist&lt;/a&gt; for that reason). AI has made convincing lures cheap, as I wrote in &lt;a href="https://kunalvohra.com/blog/can-ai-hack-systems" rel="noopener noreferrer"&gt;Can AI Hack Systems?&lt;/a&gt;, so remove whole classes of attack in configuration rather than hoping people spot better fakes.&lt;/p&gt;

&lt;p&gt;If nobody on your team owns this, it doesn't need a full-time hire. A &lt;a href="https://kunalvohra.com/technical-advisor" rel="noopener noreferrer"&gt;technical advisor&lt;/a&gt; retainer can review your tenant settings with whoever runs IT, and a &lt;a href="https://kunalvohra.com/fractional-cto" rel="noopener noreferrer"&gt;fractional CTO&lt;/a&gt; can own identity, security and the rest of the platform end to end.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;What is device code phishing?&lt;/strong&gt; It is a phishing technique that abuses the OAuth device authorization flow, which is built for TVs and command-line tools. The attacker requests a sign-in code, persuades you to type it into the real Microsoft page and sign in, and Microsoft then issues your account's tokens to the attacker's session.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Does MFA stop device code phishing?&lt;/strong&gt; Ordinary MFA does not, because you complete it yourself on the genuine Microsoft page. Blocking the device code flow with Conditional Access does, and passkeys stop the proxy-style phishing that often comes with it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How do I block device code flow in Microsoft 365?&lt;/strong&gt; Create a Conditional Access policy in the Entra admin center: all users (excluding break-glass accounts), all resources, Conditions &amp;gt; Authentication flows &amp;gt; Device code flow, grant Block access. Run it in report-only first and check sign-in logs filtered by the Device code authentication protocol. Conditional Access needs Entra ID P1.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is the OpenAI Academy event email a scam?&lt;/strong&gt; The pages in this campaign are fakes; OpenAI, SpaceX and Anthropic are being impersonated, not involved. Treat any event page that wants your work sign-in to hand out a free AI subscription as phishing, and check the company's own website directly.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Does this affect Google Workspace?&lt;/strong&gt; Yes, through a different method. The campaign's Google path is a reverse proxy that captures your password, MFA response and session cookie on a fake copy of Google's sign-in page. Passkeys and security keys defeat that, because they only work on the real Google domain.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What should I do if I entered a code from a website?&lt;/strong&gt; Tell whoever runs IT immediately. They should disable the account, revoke all sessions, reset the password, remove unknown MFA methods and devices, and check mailbox rules, app consents and file sharing before turning the account back on.&lt;/p&gt;

&lt;p&gt;If you'd like a second pair of eyes on your Microsoft or Google tenant before this lands in your inbox, &lt;a href="https://calendly.com/kunalvohra" rel="noopener noreferrer"&gt;book a 30-minute call&lt;/a&gt;. Bring an admin who can share the screen, and we'll go through the settings that matter first.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://pushsecurity.com/blog/ai-themed-fake-event-campaign-analysis" rel="noopener noreferrer"&gt;Analyzing a fake event campaign combining AiTM and device code phishing payloads&lt;/a&gt; (Push Security)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.bleepingcomputer.com/news/security/fake-openai-spacex-events-used-to-steal-microsoft-google-accounts/" rel="noopener noreferrer"&gt;Fake OpenAI, SpaceX events used to steal Microsoft, Google accounts&lt;/a&gt; (BleepingComputer)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://cybersecuritynews.com/skyw4lker-phishing-campaign-expands-threatening-us-enterprises-with-microsoft-365-account-takeovers/" rel="noopener noreferrer"&gt;SkyW4lker phishing campaign expands, threatening US enterprises with Microsoft 365 account takeovers&lt;/a&gt; (Cyber Security News)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://thehackernews.com/2026/11/weekly-recap-critical-cves-linux-ddos.html" rel="noopener noreferrer"&gt;Weekly Recap: Critical CVEs, Linux DDoS Botnet, Device-Code Phishing, Android Spyware and More&lt;/a&gt; (The Hacker News)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://thehackernews.com/2026/11/5-active-cyber-threats-that-put-us.html" rel="noopener noreferrer"&gt;5 Active Cyber Threats That Put US Enterprises at Risk in October 2026&lt;/a&gt; (The Hacker News, partner content from ANY.RUN)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.microsoft.com/en-us/security/blog/2025/02/13/storm-2372-conducts-device-code-phishing-campaign/" rel="noopener noreferrer"&gt;Storm-2372 conducts device code phishing campaign&lt;/a&gt; (Microsoft Security Blog)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://learn.microsoft.com/en-us/entra/identity/conditional-access/policy-block-authentication-flows" rel="noopener noreferrer"&gt;Block authentication flows with Conditional Access policy&lt;/a&gt; (Microsoft Learn)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://learn.microsoft.com/en-us/entra/identity/conditional-access/concept-authentication-flows" rel="noopener noreferrer"&gt;Authentication flows as a condition in Conditional Access policy&lt;/a&gt; (Microsoft Learn)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://learn.microsoft.com/en-us/entra/identity/conditional-access/concept-token-protection" rel="noopener noreferrer"&gt;Token protection in Microsoft Entra Conditional Access&lt;/a&gt; (Microsoft Learn)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://learn.microsoft.com/en-us/entra/identity/users/users-revoke-access" rel="noopener noreferrer"&gt;Revoke user access in Microsoft Entra ID&lt;/a&gt; (Microsoft Learn)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://knowledge.workspace.google.com/admin/security/deploy-2-step-verification" rel="noopener noreferrer"&gt;Deploy 2-step verification&lt;/a&gt; (Google Workspace Admin Help)&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>security</category>
      <category>cybersecurity</category>
      <category>webdev</category>
      <category>startup</category>
    </item>
    <item>
      <title>OpenAI's 948 Linux Root Exploits (November 2026): What AI-Found Kernel Bugs Mean for Your Servers</title>
      <dc:creator>Kunal Vohra</dc:creator>
      <pubDate>Sat, 07 Nov 2026 12:00:00 +0000</pubDate>
      <link>https://dev.to/kunalvohra/openais-948-linux-root-exploits-november-2026-what-ai-found-kernel-bugs-mean-for-your-servers-3obn</link>
      <guid>https://dev.to/kunalvohra/openais-948-linux-root-exploits-november-2026-what-ai-found-kernel-bugs-mean-for-your-servers-3obn</guid>
      <description>&lt;p&gt;On 4 November 2026 OpenAI published research saying an upcoming, cyber-focused model, running inside an agentic pipeline, found 948 distinct Linux kernel bugs that let an unprivileged local user become root, and wrote working exploits for them, in a single run in mid-October that cost $1,213,430 in API credits. The bugs were in current kernel code on a stock desktop install of a common distribution, with the usual hardening switched on. Most were not yet fixed ten days later. For a startup the lesson is not "panic about the kernel". It is that a foothold on your server (a compromised app process, a leaked SSH key, a malicious CI job) should now be treated as root. If your security quietly depends on "they only got the app user", fix that this week.&lt;/p&gt;

&lt;p&gt;I'm Kunal Vohra, a security researcher and cybersecurity advisor with an M.Tech in Cybersecurity from NIT Kurukshetra, and I work as a fractional CTO for AI-first startups. This post covers what OpenAI reported, where the reporting disagrees, and the practical changes I'd make to a typical startup stack.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who should care
&lt;/h2&gt;

&lt;p&gt;You should read this if any of these are true:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;You run Linux servers or VMs yourself&lt;/strong&gt; , for example a VPS with Nginx and a Node, Python or PHP app, or EC2 instances you patch by hand.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Your app user has more power than it needs&lt;/strong&gt; : it can &lt;code&gt;sudo&lt;/code&gt;, it is in the &lt;code&gt;docker&lt;/code&gt; group, or your app runs as root inside a container started with &lt;code&gt;--privileged&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;You run other people's code&lt;/strong&gt; : customer scripts, plugins, notebooks, AI agents with a code interpreter, or self-hosted CI runners that build pull requests.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Several services or customers share one host&lt;/strong&gt; and you count on Linux user separation or plain containers to keep them apart.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Kernel updates on your servers wait for "the next maintenance window"&lt;/strong&gt; that tends not to come.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you are fully on managed platforms (Vercel, Cloud Run, Lambda, a managed database), the kernel is your provider's problem, and most of this is about your CI runners and any boxes you still operate.&lt;/p&gt;

&lt;h2&gt;
  
  
  What OpenAI actually reported
&lt;/h2&gt;

&lt;p&gt;OpenAI's post is titled "A thousand fewer paths to root". I could not load the original page while writing this, so the figures below come from coverage by SecurityWeek, Help Net Security, The Hacker News, GBHackers, Decrypt and Cybersecurity News, which agree on the core numbers. Check the original before quoting it anywhere important.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The target.&lt;/strong&gt; Current Linux kernel code (SecurityWeek says the master branch), tested on an unmodified desktop install of a widely used distribution, with KASLR, SMEP and SMAP enabled. The distribution was not named. The agents audited about 7 percent of the kernel, roughly 5,034 audit targets and 2.8 million lines, chosen because unprivileged users can reach them.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The funnel.&lt;/strong&gt; Agents flagged 2,308 candidate bugs within about 12 hours. A triage stage reproduced 1,724 of them (74.7 percent) in sandboxes within about 30 hours. The pipeline produced 291 information-leak proofs of concept and 1,169 working root exploits, which grouped down to 948 unique vulnerabilities, each with a proposed patch.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The cost.&lt;/strong&gt; $1,213,430 in API credits, or about $1,280 per unique vulnerability. A July run with GPT-5.6 Sol found 166 unique privilege escalations for $513,690, so in three months the count went up almost six times while the total cost a little more than doubled.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The hardened test.&lt;/strong&gt; On Google's kernelCTF setup, which is deliberately hardened, a 112-hour run produced 21 privilege escalation exploits for about $33,000.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;New techniques.&lt;/strong&gt; The exploits used at least 10 exploitation techniques that OpenAI says outside experts had not seen before. Details were withheld.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Disclosure.&lt;/strong&gt; Each exploit and patch went through human review before being reported to kernel maintainers, starting with the most serious bugs. Ten days in, reporting says 61 bugs were patched upstream, 43 had patches in staging and 299 more were being reported or discussed. No CVE list has been published.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;OpenAI's own advice.&lt;/strong&gt; A low-privilege account on an unpatched operating system should be considered, in OpenAI's words, "functionally equivalent to root access for a well-resourced threat actor". It also points to stronger isolation such as Firecracker and gVisor.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Where the coverage differs: SecurityWeek, Decrypt, Help Net Security and GBHackers describe the model as unreleased or upcoming. Cybersecurity News reports that OpenAI announced a model called GPT-6 Cyber the same day, going first to a small group of approved customers, and links the Linux results to it. The duration also varies by outlet: 12 and 30 hours describe stages of the run, while Help Net Security says the scan and manual analysis finished after 10 days. I've used the stage figures and left the model name open.&lt;/p&gt;

&lt;p&gt;This lands one week after Anthropic's 27 October research, in which an unreleased internal model built five exploit chains, including the remote Linux kernel bug I covered in the &lt;a href="https://kunalvohra.com/blog/linux-kernel-cve-2026-98432-how-to-fix" rel="noopener noreferrer"&gt;CVE-2026-98432 fix guide&lt;/a&gt;. The two are different bug classes: Anthropic's was remote and needed no account, OpenAI's are local and need a foothold first. Together they make the same point. Finding and weaponising kernel bugs is becoming something you can buy with API credits.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why this changes the planning assumption
&lt;/h2&gt;

&lt;p&gt;For years, a lot of startup security has rested on a quiet assumption: if an attacker gets in through the app, they get the app user, and the app user can't do much. Local privilege escalation bugs always existed, but finding a fresh one took a skilled person weeks, so for most attackers the app user really was a ceiling.&lt;/p&gt;

&lt;p&gt;At around $1,300 per bug, that ceiling looks thin. You don't need to believe every attacker has this model today. OpenAI and Anthropic both limit access to their strongest cyber models. But the price curve is public now, and the people who most want this capability are the ones who will pay for it first.&lt;/p&gt;

&lt;p&gt;The line I'd put on the wall: &lt;strong&gt;a shell on your server is root, it just hasn't asked yet.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;That doesn't mean your current defences are useless. It means the layers that assume the kernel holds (Linux users, plain containers, "the CI runner only has repo access") count for less, and the layers that don't depend on it (patching speed, VM or microVM boundaries, short-lived credentials, keeping secrets off the box) count for more.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to change in your stack this week
&lt;/h2&gt;

&lt;p&gt;Eight changes, roughly in order of effort against payoff. None of them needs this research to be exactly right; all of them were good practice before it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Find out what your app user can already do.&lt;/strong&gt; If the account your app runs as can &lt;code&gt;sudo&lt;/code&gt; or talk to Docker, a kernel exploit isn't even needed. On many Ubuntu cloud images the default &lt;code&gt;ubuntu&lt;/code&gt; user has passwordless sudo, and a lot of deployment guides run the app as that user.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# Which user is the app process running as?&lt;/span&gt;
ps -eo user,pid,cmd &lt;span class="o"&gt;|&lt;/span&gt; grep -E &lt;span class="s2"&gt;"node|pm2|gunicorn|php-fpm"&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt; grep -v grep

&lt;span class="c"&gt;# Can that user sudo? (replace appuser)&lt;/span&gt;
sudo -l -U appuser

&lt;span class="c"&gt;# Is it in the docker group? Docker group membership is root-equivalent.&lt;/span&gt;
id appuser

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If the answer is "yes, it can sudo", create a dedicated, non-login user for the app, move the app and its process manager to it, and keep the admin account for humans only.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Lock the service down with systemd.&lt;/strong&gt; If the app runs under systemd (or PM2 started by systemd), a few directives remove a lot of what an attacker can reach, even before a kernel exploit is involved.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight systemd"&gt;&lt;code&gt;&lt;span class="c1"&gt;# /etc/systemd/system/myapp.service (excerpt)
&lt;/span&gt;&lt;span class="k"&gt;[Service]&lt;/span&gt;
&lt;span class="na"&gt;User&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;appuser&lt;/span&gt;
&lt;span class="na"&gt;NoNewPrivileges&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="kc"&gt;true&lt;/span&gt;
&lt;span class="na"&gt;ProtectSystem&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;strict&lt;/span&gt;
&lt;span class="na"&gt;ProtectHome&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="kc"&gt;true&lt;/span&gt;
&lt;span class="na"&gt;PrivateTmp&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="kc"&gt;true&lt;/span&gt;
&lt;span class="na"&gt;ReadWritePaths&lt;/span&gt;&lt;span class="o"&gt;=/&lt;/span&gt;&lt;span class="n"&gt;srv&lt;/span&gt;&lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="n"&gt;myapp&lt;/span&gt;&lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="n"&gt;data&lt;/span&gt;
&lt;span class="na"&gt;CapabilityBoundingSet&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;
&lt;span class="na"&gt;RestrictSUIDSGID&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="kc"&gt;true&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then run &lt;code&gt;systemd-analyze security myapp&lt;/code&gt; to see what is still exposed. These settings don't stop a kernel bug, but they make the first steps of an attack noisier and narrower.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Stop running containers as root or privileged.&lt;/strong&gt; Plain containers share the host kernel, so a kernel privilege escalation inside a container is usually a host compromise. At minimum: a non-root &lt;code&gt;USER&lt;/code&gt; in the Dockerfile, no &lt;code&gt;--privileged&lt;/code&gt;, no mounted Docker socket, dropped capabilities and Docker's default seccomp profile left on.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# Spot the risky ones on a host&lt;/span&gt;
docker ps -q &lt;span class="o"&gt;|&lt;/span&gt; xargs docker inspect \
  --format &lt;span class="s1"&gt;'{{.Name}} user={{.Config.User}} privileged={{.HostConfig.Privileged}}'&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;An empty &lt;code&gt;user=&lt;/code&gt; means the container runs as root.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Put untrusted code behind a real boundary.&lt;/strong&gt; If you run code you didn't write (customer scripts, AI agent code interpreters, plugin systems, pull-request builds from forks), a shared kernel is no longer a boundary I'd rely on. Use a separate VM per tenant or job, or a sandbox built for this such as gVisor or Firecracker microVMs, which is the direction OpenAI itself points to. This matters most for AI products: an agent that can execute code is exactly the low-privilege foothold this research is about. My &lt;a href="https://kunalvohra.com/blog/ai-agent-security" rel="noopener noreferrer"&gt;AI agent security guide&lt;/a&gt; covers the permission side.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. Make kernel patching a routine, not an event.&lt;/strong&gt; Expect more kernel fixes over the coming months as reported bugs are fixed upstream and backported. Decide how kernel updates reach your servers and how often you reboot: unattended security updates plus a scheduled reboot window, live patching (Ubuntu Livepatch, kpatch on RHEL) for hosts you can't restart often, or managed node pools with automatic upgrades on Kubernetes. Then check whether a reboot is pending:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;uname -r &lt;span class="c"&gt;# running kernel&lt;/span&gt;
ls /var/run/reboot-required 2&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;/dev/null &lt;span class="c"&gt;# Ubuntu/Debian: a reboot is pending&lt;/span&gt;
needrestart -k 2&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;/dev/null &lt;span class="c"&gt;# if installed: running vs installed kernel&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Patch the package manager first; the &lt;a href="https://kunalvohra.com/blog/apt-cve-2026-109518-how-to-fix" rel="noopener noreferrer"&gt;APT CVE-2026-109518 guide&lt;/a&gt; explains why that order mattered this month.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;6. Treat CI runners as hostile.&lt;/strong&gt; Self-hosted runners that build pull requests, especially from forks, are a gift to anyone with a local exploit. Use ephemeral runners that are destroyed after each job, never share a runner between public and private repositories, and keep deploy credentials out of jobs that build untrusted code.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;7. Assume root reads everything on the box.&lt;/strong&gt; If an attacker reaches root, every &lt;code&gt;.env&lt;/code&gt; file, cloud credential and SSH key on that host is theirs. Move long-lived keys to a secrets manager, prefer short-lived cloud credentials through instance roles or workload identity, and scope each key to one service so one host can't unlock the rest.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;8. Watch for the escalation, not just the entry.&lt;/strong&gt; A web shell running as the app user and then a root process appearing from it is a pattern you can alert on. auditd rules or a runtime tool such as Falco can flag new setuid binaries, unexpected root processes whose parent is your app, and kernel module loads. If you have nothing, start by shipping auth and sudo logs off the box.&lt;/p&gt;

&lt;p&gt;If you only have an hour, do steps 1, 3 and 5. They close the most common gaps I see in startup infrastructure reviews, and they are on my &lt;a href="https://kunalvohra.com/blog/startup-security-checklist" rel="noopener noreferrer"&gt;startup security checklist&lt;/a&gt; for that reason.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I'd do as a CTO
&lt;/h2&gt;

&lt;p&gt;I'd book one engineer-day this week. In the morning, inventory every host and runner we operate, and for each one answer three questions: what runs as which user, can that user become root without an exploit, and when was the kernel last updated. In the afternoon, fix whatever came back "yes" or "months ago", starting with anything that runs untrusted code or holds production credentials.&lt;/p&gt;

&lt;p&gt;I'd also change one line in our threat model. "Attacker gets the app user" moves from a contained incident to "attacker gets the host". That changes how we rate application bugs too: a file upload flaw or a server-side request forgery that gives code execution as the app user is now a host compromise, and it goes to the top of the queue.&lt;/p&gt;

&lt;p&gt;And I'd point the same technology at our own code. Anthropic's advice after its own research was blunt: scanning your own code matters more than ever. Use the AI code review you already pay for on pull requests, ask it specifically about authentication, file handling and command execution, and keep a human deciding what ships. Defenders have the same tools now; what matters is who uses them first.&lt;/p&gt;

&lt;p&gt;If nobody on your team owns this kind of decision, it doesn't need a full-time hire. A &lt;a href="https://kunalvohra.com/technical-advisor" rel="noopener noreferrer"&gt;technical advisor&lt;/a&gt; retainer can review the host inventory and the fixes with your engineers, and a &lt;a href="https://kunalvohra.com/fractional-cto" rel="noopener noreferrer"&gt;fractional CTO&lt;/a&gt; can own security, infrastructure and the patching routine end to end. For the broader picture of AI-driven attacks, see &lt;a href="https://kunalvohra.com/blog/can-ai-hack-systems" rel="noopener noreferrer"&gt;Can AI Hack Systems?&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Are the 948 Linux bugs zero-days?&lt;/strong&gt; They are newly found bugs in current kernel code, reported privately to the kernel maintainers. OpenAI has not published a list of them or CVE numbers. About ten days after the run, reporting says 61 were patched upstream and 43 had staged patches, so most were still unfixed at that point. Exploit details have not been released.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is my Linux server affected?&lt;/strong&gt; Probably, in the sense that these are kernel bugs and the kernel is shared by almost every Linux system. But every bug here needs local code execution first. The practical question is whether an attacker can get a foothold (through your app, an exposed service, a stolen key or a CI job) and what that foothold can already do. Fix that, and keep kernels updated as backported fixes arrive.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Can attackers use OpenAI's model to do this?&lt;/strong&gt; Not openly today, according to the coverage: OpenAI describes the model as upcoming, and its cyber models are offered to vetted customers under access controls. One outlet reports a GPT-6 Cyber model launching to a small approved group. Plan on the assumption that similar capability will reach well-funded attackers, because the cost per bug is already low.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Does running my app in Docker protect me?&lt;/strong&gt; Less than people think. Containers share the host kernel, so a kernel privilege escalation inside a container usually means the host. Non-root users, no privileged mode, dropped capabilities and seccomp help a lot. For untrusted code, use VM or microVM isolation such as gVisor or Firecracker.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What is local privilege escalation?&lt;/strong&gt; It is a bug that lets someone who can already run code on a machine as an ordinary user gain more rights, usually root. On its own it doesn't let an attacker in from the internet; it turns a small foothold into full control.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Should my startup use AI to find bugs in our own code?&lt;/strong&gt; Yes, with human review. Running AI-assisted review on every pull request and periodically re-scanning older code is cheap compared with an incident, and it is the defensive side of the same shift. Treat its findings as leads to verify, not verdicts.&lt;/p&gt;

&lt;p&gt;If you want a second pair of eyes on your servers, containers or CI before the next round of kernel fixes lands, &lt;a href="https://calendly.com/kunalvohra" rel="noopener noreferrer"&gt;book a 30-minute call&lt;/a&gt;. Bring your host list and how you deploy, and we'll find the gaps that matter first.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://openai.com/index/a-thousand-fewer-paths-to-root/" rel="noopener noreferrer"&gt;A thousand fewer paths to root&lt;/a&gt; (OpenAI)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.securityweek.com/ai-agents-got-root-on-linux-via-948-kernel-flaws-in-openai-research/" rel="noopener noreferrer"&gt;AI agents got root on Linux via 948 kernel flaws in OpenAI research&lt;/a&gt; (SecurityWeek)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.helpnetsecurity.com/2026/11/05/openai-linux-kernel-root-exploits/" rel="noopener noreferrer"&gt;OpenAI's upcoming AI model built 948 working Linux root exploits in a single run&lt;/a&gt; (Help Net Security)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://thehackernews.com/2026/11/openai-ai-pipeline-generates-948-linux.html" rel="noopener noreferrer"&gt;OpenAI AI pipeline generates 948 Linux root escalation exploits in 30-hour run&lt;/a&gt; (The Hacker News)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://gbhackers.com/openais-new-cyber-ai-model-finds-948-linux-privilege-escalation-vulnerabilities/" rel="noopener noreferrer"&gt;OpenAI's new cyber AI model finds 948 Linux privilege escalation vulnerabilities&lt;/a&gt; (GBHackers)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://decrypt.co/382632/openai-linux-security-flaws-exploit" rel="noopener noreferrer"&gt;OpenAI found 948 Linux security flaws and figured out how to exploit them&lt;/a&gt; (Decrypt)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://cybersecuritynews.com/openai-launches-gpt-6-cyber/" rel="noopener noreferrer"&gt;OpenAI launches GPT-6 Cyber&lt;/a&gt; (Cybersecurity News)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.anthropic.com/research/discovering-vulnerabilities-with-claude" rel="noopener noreferrer"&gt;Discovering and exploiting vulnerabilities in critical, real-world software with Claude&lt;/a&gt; (Anthropic)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://thehackernews.com/2026/10/anthropics-unreleased-ai-model-chains.html" rel="noopener noreferrer"&gt;Anthropic's unreleased AI model chains exploits&lt;/a&gt; (The Hacker News)&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>programming</category>
      <category>machinelearning</category>
      <category>startup</category>
    </item>
    <item>
      <title>Technical Debt in Startups: How to Manage It, Prioritise It, and Pay It Down Without Stopping to Rewrite (2026)</title>
      <dc:creator>Kunal Vohra</dc:creator>
      <pubDate>Fri, 06 Nov 2026 12:00:00 +0000</pubDate>
      <link>https://dev.to/kunalvohra/technical-debt-in-startups-how-to-manage-it-prioritise-it-and-pay-it-down-without-stopping-to-2904</link>
      <guid>https://dev.to/kunalvohra/technical-debt-in-startups-how-to-manage-it-prioritise-it-and-pay-it-down-without-stopping-to-2904</guid>
      <description>&lt;p&gt;Technical debt in startups is not a failure; it is borrowing. Every startup that ships fast takes on some, and the ones that ship without any usually shipped too late. The problem is never that debt exists. The problem is debt nobody has listed, nobody has priced, and nobody owns, which is how a team that moved fast in month three is crawling in month twelve.&lt;/p&gt;

&lt;p&gt;I'm Kunal Vohra. I've been the technical co-founder six times, and as a fractional and interim CTO a large part of my job is walking into a codebase that has run up debt and deciding, with the founder, what to pay down and what to live with. This is the method I use: how to sort technical debt, how to prioritise it, how much time to give it, and who should own it at your stage.&lt;/p&gt;

&lt;p&gt;If you have inherited a product from an agency and are already asking whether to start again, read my guide on &lt;a href="https://kunalvohra.com/blog/rebuild-or-refactor-mvp" rel="noopener noreferrer"&gt;whether to rebuild or refactor an MVP&lt;/a&gt; first. This post is about the ongoing version of the problem: a product you are still building, with a team you still have.&lt;/p&gt;

&lt;h2&gt;
  
  
  What technical debt in startups actually is
&lt;/h2&gt;

&lt;p&gt;Technical debt is every shortcut that makes today faster and some future change slower, riskier or more expensive. Some of it is deliberate (no automated tests on a feature you might delete next month), some is accidental (a data model designed before you understood the business), and some simply happens as the product outgrows decisions that were right at the time.&lt;/p&gt;

&lt;p&gt;Founders tend to hear "tech debt" as "the engineers want to clean up". That framing is why it gets ignored until it hurts. The useful framing is financial: debt has a principal (what it would cost to fix) and an interest rate (what it costs you every week you don't). A shortcut with no interest is free money. A shortcut with high interest compounds quietly until every release is slow and every estimate doubles.&lt;/p&gt;

&lt;p&gt;The signs that the interest has started compounding are consistent:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Small changes take weeks.&lt;/strong&gt; A feature that should be a day becomes a sprint, and nobody can say exactly why.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Fixing one thing breaks another.&lt;/strong&gt; Releases cause regressions in parts of the product nobody touched.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Only one person can touch certain areas.&lt;/strong&gt; Knowledge, not code, has become the bottleneck.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Deploys are events.&lt;/strong&gt; Releasing needs a ritual, a specific person, or a quiet Friday.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Onboarding a new engineer takes a month.&lt;/strong&gt; The codebase can't be learned, only absorbed.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If several of these are true, the debt is no longer a background cost. It is setting your roadmap.&lt;/p&gt;

&lt;h2&gt;
  
  
  Good vs bad technical debt: the three buckets I use
&lt;/h2&gt;

&lt;p&gt;Not all technical debt is the same, and treating it as one undifferentiated problem is the most common mistake I see. I sort every item into one of three buckets, each with a rough cost to fix, so the founder can decide instead of guess:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Debt that is fine forever.&lt;/strong&gt; Ugly code in a corner nobody changes, an admin screen that works but looks dated, an internal script held together with tape. It is irritating, and it costs you nothing. Leave it alone. Engineers will want to tidy it; the honest answer is "not this quarter, maybe never".&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Debt that is costing you velocity.&lt;/strong&gt; The module every feature has to touch and nobody wants to. The missing test suite on the payments path that makes every release a gamble. The tangled data access that turns a simple report into a week. This is the bucket that silently sets your release pace, and it is where most paydown effort should go.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Debt that is one failure away from an outage or a breach.&lt;/strong&gt; No tested backups. Secrets in the repository. A single server with no failover. Access control that is checked in some places and not others. This bucket is small, and it is the only one that is urgent regardless of the roadmap. My &lt;a href="https://kunalvohra.com/blog/startup-security-checklist" rel="noopener noreferrer"&gt;startup security checklist&lt;/a&gt; covers most of what lands here.&lt;/p&gt;

&lt;p&gt;Bucket three is where I am least relaxed, for a reason I learned the hard way. At the edtech venture I wrote about in &lt;a href="https://kunalvohra.com/blog/the-edtech-problem-nobody-would-touch" rel="noopener noreferrer"&gt;this case study&lt;/a&gt;, our first full run of 500,000 assessments in an hour nearly came undone because of one wrong line in an AWS load balancer configuration, caught at the last moment. &lt;strong&gt;The debt that ends companies is rarely the ugly code; it is the one line nobody checked.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  How to prioritise technical debt in a startup
&lt;/h2&gt;

&lt;p&gt;Prioritise technical debt by what it blocks, not by how bad the code looks. Every item on the list should answer one question: what does this stop us from doing, and what would fixing it unlock? Ranked that way, the list usually collapses to two or three items that matter now.&lt;/p&gt;

&lt;p&gt;Here is the order I work in:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Anything in bucket three goes first,&lt;/strong&gt; in small fixes, this cycle. Not because it is slowing you down, but because the downside is not recoverable.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Debt blocking revenue comes next.&lt;/strong&gt; An enterprise prospect needs single sign-on, audit logs or a security questionnaire answered, and the code can't support it cleanly. That is debt with a dollar figure attached.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Debt blocking releases is third.&lt;/strong&gt; Fragile deploys, no staging environment, no tests on the paths that move money or data. Fixing these makes every later fix cheaper, so they pay back fastest.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Debt slowing the roadmap is fourth.&lt;/strong&gt; The module that turns days into weeks. Fix it when the next feature on the roadmap has to go through it anyway, so the paydown rides along with work you were doing regardless.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Everything else waits.&lt;/strong&gt; Write it down so it is not forgotten, and stop discussing it.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The trick in step four matters more than it sounds. The cheapest time to pay down a piece of debt is when you are already changing that code for a feature. A standalone "refactoring sprint" competes with the roadmap and usually loses; a refactor attached to a feature the founder already wants ships.&lt;/p&gt;

&lt;h2&gt;
  
  
  How much time should a startup spend on technical debt?
&lt;/h2&gt;

&lt;p&gt;There is no universal percentage, but there should be a deliberate one, agreed between the founder and whoever leads engineering. A common rule of thumb is to reserve a fixed slice of every cycle for paydown, and many teams land somewhere around 10 to 20 percent. The exact number matters less than the fact that it exists and is protected.&lt;/p&gt;

&lt;p&gt;How I set it in practice:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Before product-market fit,&lt;/strong&gt; keep paydown minimal and targeted. Bucket three gets fixed; almost everything else waits, because the product may change shape entirely. Speed of learning beats code quality at this stage.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;After you have paying customers who depend on the product,&lt;/strong&gt; reserve a standing slice every cycle and spend it on bucket two, in the order above. Reliability is now part of what you sell.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Before a raise or a big enterprise deal,&lt;/strong&gt; spend a short, focused stretch on whatever a diligence reviewer or security questionnaire will find first. My guide to &lt;a href="https://kunalvohra.com/blog/ai-startup-technical-due-diligence" rel="noopener noreferrer"&gt;AI startup technical due diligence&lt;/a&gt; lists what investors actually check.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;What I don't recommend is the "feature freeze to clean everything up". It feels responsible, it stalls the company, and it rarely finishes, because the list was never prioritised in the first place. Small, continuous paydown beats a big one almost every time, for the same reason refactoring beats rewriting.&lt;/p&gt;

&lt;h2&gt;
  
  
  AI-written code and technical debt
&lt;/h2&gt;

&lt;p&gt;AI coding tools make you faster and make debt cheaper to create, which means it piles up faster than it used to. I use these tools every day; most of the code I ship now starts as generated code. But teams shipping mostly-generated code tend to hit a wall around month six: nobody fully understands the codebase, the tests exist but don't really test anything, and every change breaks something adjacent.&lt;/p&gt;

&lt;p&gt;The cause is not the tools; it is shipping code nobody has read. The controls that keep it in check are not complicated:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Every generated change gets a human review&lt;/strong&gt; by someone who could have written it, with the same bar as hand-written code.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Architectural boundaries are set by a person,&lt;/strong&gt; written down, and enforced in review, so generated code fills in the structure instead of inventing one each time.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Tests are checked for what they assert,&lt;/strong&gt; not just that they pass. A generated test that asserts nothing is worse than no test, because it gives false confidence.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Duplication gets watched.&lt;/strong&gt; Generators happily write the same logic five times in five places; that is bucket two debt in the making.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you are building an AI product, there is a second layer: prompts, model choices and provider integrations are code too, and they accumulate their own debt. Hard-wiring one provider everywhere is the classic example. My notes on &lt;a href="https://kunalvohra.com/fractional-cto/ai-startups" rel="noopener noreferrer"&gt;fractional CTO work for AI startups&lt;/a&gt; cover the boundaries I put in place there.&lt;/p&gt;

&lt;h2&gt;
  
  
  How to explain technical debt to your board and investors
&lt;/h2&gt;

&lt;p&gt;Explain technical debt in the board's language: risk and speed, with numbers attached. "The code is messy" gets a shrug or a panic; neither helps. A one-page debt register gets a decision.&lt;/p&gt;

&lt;p&gt;The register I write has one line per item that matters, with four things on each line:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;What it is,&lt;/strong&gt; in one plain sentence.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Which bucket it sits in&lt;/strong&gt; (fine, velocity, or outage/breach risk).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What it costs now,&lt;/strong&gt; in terms the board cares about: slower releases, a blocked deal, a risk of downtime.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What fixing it costs,&lt;/strong&gt; as a rough range of time, and what it unlocks.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Presented that way, debt stops being an engineering complaint and becomes an investment decision, which is what it is. It also protects you in diligence. Investors don't expect a seed-stage codebase to be clean; they expect the team to know where the problems are. A founder who can hand over a prioritised list looks in control. A founder who is surprised by the reviewer's findings does not.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who should own technical debt at your stage
&lt;/h2&gt;

&lt;p&gt;Technical debt needs an owner with the authority to say no to the roadmap sometimes, and the judgement to know when. At most early-stage companies that person doesn't exist yet, which is exactly why debt accumulates unmanaged. What that owner looks like depends on your stage; my guide to &lt;a href="https://kunalvohra.com/blog/what-does-a-cto-do-in-a-startup" rel="noopener noreferrer"&gt;what a CTO does in a startup&lt;/a&gt; goes through the role in detail.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;No technical leader, small team or a vendor building:&lt;/strong&gt; you need judgement more than hours. A &lt;a href="https://kunalvohra.com/technical-advisor" rel="noopener noreferrer"&gt;technical advisor&lt;/a&gt; at $750 a month for 15+ hours (held through December 2026) can sort the list into buckets, sanity-check estimates, and tell you which items actually matter.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A team that needs someone accountable:&lt;/strong&gt; a &lt;a href="https://kunalvohra.com/fractional-cto" rel="noopener noreferrer"&gt;fractional CTO&lt;/a&gt; owns the register, the paydown slice, and the trade-offs against the roadmap, usually two days a week at $2,500 a month for 50+ hours. The &lt;a href="https://kunalvohra.com/fractional-cto-cost" rel="noopener noreferrer"&gt;fractional CTO cost guide&lt;/a&gt; puts market rates beside mine.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Debt has become a crisis:&lt;/strong&gt; releases have stalled, the last technical leader has left, or a raise is close. An &lt;a href="https://kunalvohra.com/interim-cto" rel="noopener noreferrer"&gt;interim CTO&lt;/a&gt; works close to full-time for three to nine months, pays down the two or three debts costing the most, and hands over to a permanent hire. The Embedded tier is $4,000 a month for 80+ hours.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;For Series A companies this is often the whole engagement: triaging debt into those three buckets with real cost estimates, then bridging to the permanent CTO. My page on &lt;a href="https://kunalvohra.com/fractional-cto/series-a" rel="noopener noreferrer"&gt;fractional CTO for Series A startups&lt;/a&gt; describes how that runs. You rarely need a full-time CTO at $180,000 to $260,000 a year plus equity just to get debt under control; you need someone senior for the months when the calls are most expensive.&lt;/p&gt;

&lt;p&gt;And if you are still before v1, the cheapest technical debt is the debt you never take on. Choosing a boring, hireable stack (&lt;a href="https://kunalvohra.com/blog/best-tech-stack-for-mvp" rel="noopener noreferrer"&gt;best tech stack for an MVP&lt;/a&gt;) and a team that hands over clean code is most of it. My team at Panicle Tech builds MVPs that way (&lt;a href="https://kunalvohra.com/mvp-development" rel="noopener noreferrer"&gt;MVP development&lt;/a&gt;, from $30 a screen), and the &lt;a href="https://kunalvohra.com/mvp-cost-calculator" rel="noopener noreferrer"&gt;MVP cost calculator&lt;/a&gt; gives you a range before anyone quotes.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Is technical debt bad for a startup?&lt;/strong&gt; No, not by itself. Technical debt in startups is a normal consequence of shipping fast, and some of it is the right call. It becomes bad when it is unlisted and unowned, and when the "interest" (slower releases, regressions, outage risk) starts setting your roadmap.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How do you prioritise technical debt?&lt;/strong&gt; By what each item blocks. Fix anything that risks an outage or a breach first, then debt blocking revenue, then debt blocking releases, then debt slowing the roadmap (ideally while you are changing that code for a feature anyway). Everything else gets written down and left alone.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How much time should engineers spend on technical debt?&lt;/strong&gt; A deliberate, protected slice of every cycle, agreed with the founder. Many teams use roughly 10 to 20 percent as a rule of thumb. Before product-market fit keep it minimal and focused on security and reliability; after you have customers who depend on the product, make it a standing allocation.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Should we stop building features to pay down technical debt?&lt;/strong&gt; Rarely. A feature freeze to clean up usually stalls the company and doesn't finish. Small, continuous paydown attached to roadmap work beats a big cleanup. The exception is a short, focused push on outage or security risks, or before a raise or a big enterprise deal.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Does AI-generated code create more technical debt?&lt;/strong&gt; It creates debt faster, because it makes code cheap to produce. The fix is review, not avoidance: a human who could have written it reviews every change, architecture is set and enforced by a person, and tests are checked for what they actually assert.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Do I need a CTO to manage technical debt?&lt;/strong&gt; You need someone senior to own it, but rarely a full-time hire at the early stage. A technical advisor can triage and prioritise; a fractional CTO can own the register and the trade-offs against the roadmap; an interim CTO fits when debt has turned into a crisis.&lt;/p&gt;

&lt;p&gt;If your releases have slowed down and you can't tell which debts are worth paying, &lt;a href="https://calendly.com/kunalvohra" rel="noopener noreferrer"&gt;book a 30-minute call&lt;/a&gt;. Bring your team's list if there is one, or a description of what keeps breaking. I'll tell you which bucket each item belongs in and what I'd fix first. Or &lt;a href="https://kunalvohra.com/contact" rel="noopener noreferrer"&gt;get in touch&lt;/a&gt; if you'd rather write first.&lt;/p&gt;

</description>
      <category>startup</category>
      <category>career</category>
      <category>leadership</category>
      <category>entrepreneurship</category>
    </item>
    <item>
      <title>Kimi K3.1 Pro Open Weights (November 2026): Should Your Startup Use It, Self-Host It, or Skip It?</title>
      <dc:creator>Kunal Vohra</dc:creator>
      <pubDate>Fri, 06 Nov 2026 12:00:00 +0000</pubDate>
      <link>https://dev.to/kunalvohra/kimi-k31-pro-open-weights-november-2026-should-your-startup-use-it-self-host-it-or-skip-it-4301</link>
      <guid>https://dev.to/kunalvohra/kimi-k31-pro-open-weights-november-2026-should-your-startup-use-it-self-host-it-or-skip-it-4301</guid>
      <description>&lt;p&gt;Moonshot AI has released the full open weights of Kimi K3.1 Pro on Hugging Face, two weeks after launching it in the Kimi app and API on 23 October 2026. It is a 2.8-trillion-parameter mixture-of-experts model with 104 billion parameters active per token, a context window of just over one million tokens, text and image input, and API pricing of $2 per million input tokens and $10 per million output. Moonshot's own benchmarks put it within a point of Claude Opus 5.5 on agentic coding. For a startup, the practical answer is short: you can try it through an API this week, you almost certainly cannot self-host it, and you should not switch production traffic until it passes your own eval set.&lt;/p&gt;

&lt;p&gt;I'm Kunal Vohra. I run ColadAI, a multi-LLM platform, and I've been the technical co-founder six times, so "should we move to the new model?" is a question I answer every few weeks. This post is the version I'd want before a team meeting: what shipped, what it costs, what the licence says, and what to do about it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who should care about Kimi K3.1 Pro
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;AI startups paying frontier-model bills.&lt;/strong&gt; If a large share of your spend goes to a top-tier closed model for coding agents, long-document work or reasoning, a model priced at half of Claude Opus 5.5's list rate with comparable claimed scores is worth a test.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Teams that need long context.&lt;/strong&gt; One million tokens of context is enough to put a mid-sized codebase or a stack of contracts in a single request. Few open models offer that.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Founders who were told "it's open source, so we can run it ourselves."&lt;/strong&gt; The weights are open. The hardware bill is not small. More on that below.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Anyone with data-residency or vendor-risk rules.&lt;/strong&gt; Open weights mean the model can run with a host of your choice, not only Moonshot.&lt;/p&gt;

&lt;p&gt;If you are still choosing your first model, this changes little: pick a sensible default behind a thin provider layer. My &lt;a href="https://kunalvohra.com/blog/gemini-vs-claude-vs-grok-vs-chatgpt-api-for-business" rel="noopener noreferrer"&gt;comparison of the major AI APIs&lt;/a&gt; covers that setup.&lt;/p&gt;

&lt;h2&gt;
  
  
  What actually shipped, with dates
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;23 October 2026:&lt;/strong&gt; Moonshot announced Kimi K3.1 Pro and made it live in kimi.ai, the Kimi app, Kimi Code and the Kimi API. Its launch blog promised the full weights "by Nov 4".&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4 to 5 November 2026:&lt;/strong&gt; the weights appeared on Hugging Face under moonshotai/Kimi-K3.1-Pro. Reports differ on the exact day (one says 4 November, another 5 November with a timestamp that is 4 November in UTC if it was Beijing time), so treat it as "this week".&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The model, per Moonshot's model card:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Architecture:&lt;/strong&gt; mixture-of-experts, 2.8T total parameters, 104B activated, 93 layers, 896 experts with 16 selected per token plus 2 shared.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Attention:&lt;/strong&gt; a hybrid of KDA and Gated DSA layers, which is Moonshot's route to the long context without the memory cost of full attention everywhere.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Context:&lt;/strong&gt; 1,048,576 tokens.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Input:&lt;/strong&gt; text and images, through a 401M-parameter vision encoder (MoonViT-V2).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Precision:&lt;/strong&gt; Moonshot says MXFP4 weights and MXFP8 activations with quantization-aware training. The Hugging Face metadata tags the files as 8-bit compressed tensors, so check the exact format your inference engine expects.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Download size:&lt;/strong&gt; 1.56 TB across 96 safetensors shards.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Inference engines named on the card:&lt;/strong&gt; vLLM, SGLang and TokenSpeed.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Pricing on Moonshot's API (USD per million tokens):&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;K3.1 Pro:&lt;/strong&gt; $2.00 input, $0.20 cached input, $10.00 output.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;K3.1 Pro Highspeed:&lt;/strong&gt; $4.00 input, $0.20 cached input, $20.00 output, which Moonshot says runs two to three times faster.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;K3 (the previous model):&lt;/strong&gt; $3.00 input, $15.00 output. The new model is cheaper than the old one.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Benchmarks, all as reported by Moonshot.&lt;/strong&gt; On Terminal-Bench 4.0 it scores 65.7 against 66.4 for Claude Opus 5.5. On DeepSWE v1.1 it scores 75.5 against 74.2 for Opus 5.5, though Moonshot's own table shows DeepSeek-V4.1-Flash higher at 77.0. On SWE-Marathon it reports 51.7 against 46.3. It trails on some reasoning sets, for example 39.4 on HLE-Diamond against 60.9 for Opus 5.5. These are the vendor's numbers, run with its own settings (maximum reasoning effort, mostly inside Claude Code as the agent harness).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Independent numbers disagree.&lt;/strong&gt; Artificial Analysis lists an Intelligence Index of 46 on its v5.0.1 scale. Other write-ups quote 55, which may be an older version of the index. I'd use the Artificial Analysis page itself and treat any single index as a rough guide, not a verdict.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The safety claim.&lt;/strong&gt; Moonshot says it removed cyber-offence data from both the mid-training and reinforcement-learning stages, and that the open model's hacking ability remains behind frontier closed models that have been public for at least three months. Its own internal exploit benchmark shows a drop (81.0 with the removal against 90.5 without). That is a meaningful, documented step. It is also a vendor claim that independent red teams have not yet tested in public, and it arrives the same week Acronis described a single attacker running a mesh of AI agents on jailbroken open-weight models against organisations in 11 countries.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Open weights are a licence to choose where a model runs. They are not a discount on the hardware it needs.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  What the licence allows
&lt;/h2&gt;

&lt;p&gt;Kimi K3.1 Pro ships under its own "Kimi K3.1 Pro License", not Apache 2.0 or MIT. It grants broad rights (use, modify, fine-tune, deploy, distribute, sell) with conditions. The three that matter:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Keep the notice.&lt;/strong&gt; Include the copyright and licence text with any copy.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Model-as-a-service above $20 million.&lt;/strong&gt; If you or an affiliate run a business selling access to the model itself, and that business makes more than $20 million in revenue over any 12 consecutive months, you need a separate agreement with Moonshot before commercial use. Building features on top of the model inside your own product is not model-as-a-service under the licence.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Attribution at very large scale.&lt;/strong&gt; A commercial product with more than 100 million monthly active users, or more than $20 million in monthly revenue, must display "Kimi K3.1-Pro" prominently in its interface.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Internal use that does not expose the model or its outputs to third parties is exempt, as is use through Moonshot's own products or certified inference partners.&lt;/p&gt;

&lt;p&gt;For nearly every startup reading this, none of the thresholds bite. The one to watch is if you are building an inference platform, a model router or a "bring your own model" product: that is close to the model-as-a-service definition, and you should get a lawyer to read section 2 before you scale. I'm not a lawyer, and licences change between versions, so read the file attached to the exact checkpoint you deploy.&lt;/p&gt;

&lt;h2&gt;
  
  
  Can you self-host Kimi K3.1 Pro? Almost certainly not
&lt;/h2&gt;

&lt;p&gt;The weights alone are 1.56 TB. Moonshot publishes no official hardware configuration, and the estimates I've seen from hosting providers start at more than 1.56 TB of combined GPU memory and go to 2 TB or more once you leave room for the memory each conversation needs.&lt;/p&gt;

&lt;p&gt;Some rough arithmetic to make that concrete. An H100 has 80 GB of memory. Holding 1.56 TB of weights takes about 20 of them before a single user's request is served, and a realistic deployment would need more, spread across several eight-GPU nodes with fast interconnect. In my &lt;a href="https://kunalvohra.com/blog/self-hosting-llm-vs-api-cost" rel="noopener noreferrer"&gt;self-hosting break-even post&lt;/a&gt; I used a median H100 rental of $3.81 an hour from October 2026. Twenty GPUs at that rate, running all month, is roughly $55,600 a month, and that figure is the floor, not the plan. It leaves out the extra GPUs for context, the second copy for failover, and the engineers to keep it alive.&lt;/p&gt;

&lt;p&gt;At $2 / $10 per million tokens, $55,600 buys roughly 2.1 million requests a month at 8,000 tokens in and 1,000 out. Very few startups send that much traffic to one frontier-class model. So the self-hosting question for this model is closed for almost everyone, the same way it was for the 500B-class models in my earlier post.&lt;/p&gt;

&lt;p&gt;What open weights do give you is choice of host. Moonshot's partner list for the previous model, K3, named Together AI, Fireworks, AWS Bedrock and DeepInfra, and the Hugging Face page shows Together AI as an inference provider for K3.1 Pro. I could not confirm a full K3.1 Pro price list on any third-party host at the time of writing. Check before you plan around one.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it costs per request, compared with Claude Opus 5.5
&lt;/h2&gt;

&lt;p&gt;Using the same request shape as my other cost posts (a retrieval-heavy turn with 8,000 tokens in and 1,000 out):&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Kimi K3.1 Pro:&lt;/strong&gt; 8,000 × $2 + 1,000 × $10, per million, is about $0.026 per request.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Kimi K3.1 Pro with 6,000 of those input tokens cached:&lt;/strong&gt; about $0.015.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Kimi K3.1 Pro Highspeed:&lt;/strong&gt; about $0.052.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Claude Opus 5.5&lt;/strong&gt; at its October list price of $4 / $20: about $0.052.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;So the standard tier costs half of Opus 5.5 per request, at the same price as the mid-tier closed models (Claude Sonnet 5.5 and GPT-6.1 Sol were both $2 / $10 in October). The Highspeed tier costs the same as Opus 5.5. The honest framing is not "frontier quality at half price". It is "a new option at mid-tier prices that claims near-frontier scores on coding agents". Whether that holds on your workload is the only question that matters.&lt;/p&gt;

&lt;p&gt;Two costs hide outside the token price: reasoning at maximum effort produces many output tokens (Artificial Analysis measured about $3.68 and 36 minutes per task on its suite), and prompts, tool schemas and parsers tuned for one model often need rework for another.&lt;/p&gt;

&lt;h2&gt;
  
  
  A one-week evaluation plan before you switch
&lt;/h2&gt;

&lt;p&gt;This is the sequence I'd run with a team: about a week of part-time effort, ending in a decision with numbers behind it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Day 1: decide what you are testing for.&lt;/strong&gt; Pick one workload where cost or quality hurts today: your coding agent, long-document extraction, or support answers. Write down the current model, its cost per task and its pass rate. If you don't log cost per request yet, start there; my &lt;a href="https://kunalvohra.com/blog/ai-mvp-cost" rel="noopener noreferrer"&gt;AI MVP cost guide&lt;/a&gt; shows the per-user arithmetic.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Day 1: check the data path.&lt;/strong&gt; Read Moonshot's API data-retention and training terms, or those of whichever host you use, against your customer contracts. If your contracts name approved sub-processors, a new provider may need a customer notice before any production data touches it. Test with synthetic or scrubbed data until that is settled.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Days 2 to 3: build or reuse an eval set.&lt;/strong&gt; Fifty to two hundred real tasks with known good answers or a clear grading rubric. Reuse an existing one unchanged so the comparison is fair.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Days 3 to 4: run both models through the same harness.&lt;/strong&gt; Keep the provider behind one interface so only the configuration changes. A minimal shape, assuming an OpenAI-compatible endpoint (take the base URL and model id from your provider's docs, not from this post):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;time&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;json&lt;/span&gt;
&lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="n"&gt;openai&lt;/span&gt; &lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="nc"&gt;OpenAI&lt;/span&gt;

&lt;span class="no"&gt;MODELS&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="s2"&gt;"current"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="s2"&gt;"base_url"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;environ&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"CURRENT_BASE_URL"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                &lt;span class="s2"&gt;"api_key"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;environ&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"CURRENT_API_KEY"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                &lt;span class="s2"&gt;"model"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;environ&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"CURRENT_MODEL"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="s2"&gt;"candidate"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="s2"&gt;"base_url"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;environ&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"CANDIDATE_BASE_URL"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                  &lt;span class="s2"&gt;"api_key"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;environ&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"CANDIDATE_API_KEY"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                  &lt;span class="s2"&gt;"model"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;environ&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"CANDIDATE_MODEL"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;

&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;run&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;name&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;task&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;cfg&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="no"&gt;MODELS&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;name&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
    &lt;span class="n"&gt;client&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;OpenAI&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;base_url&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;cfg&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"base_url"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;api_key&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;cfg&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"api_key"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="n"&gt;start&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;time&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;time&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="n"&gt;resp&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;client&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;chat&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;completions&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;create&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
        &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;cfg&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"model"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="n"&gt;messages&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="s2"&gt;"role"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"user"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"content"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;task&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"prompt"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="s2"&gt;"model"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;name&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="s2"&gt;"task_id"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;task&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"id"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="s2"&gt;"output"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;resp&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;choices&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;message&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;content&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="s2"&gt;"input_tokens"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;resp&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;usage&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;prompt_tokens&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="s2"&gt;"output_tokens"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;resp&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;usage&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;completion_tokens&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="s2"&gt;"seconds"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nb"&gt;round&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;time&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;time&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;start&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;

&lt;span class="k"&gt;with&lt;/span&gt; &lt;span class="nb"&gt;open&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"eval_tasks.jsonl"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;as&lt;/span&gt; &lt;span class="n"&gt;f&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;tasks&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;loads&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;line&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;line&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="n"&gt;f&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;

&lt;span class="k"&gt;with&lt;/span&gt; &lt;span class="nb"&gt;open&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"results.jsonl"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"w"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;as&lt;/span&gt; &lt;span class="n"&gt;out&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;task&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="n"&gt;tasks&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;name&lt;/span&gt; &lt;span class="k"&gt;in&lt;/span&gt; &lt;span class="no"&gt;MODELS&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
            &lt;span class="n"&gt;out&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;write&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;json&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;dumps&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nf"&gt;run&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;name&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;task&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="s2"&gt;"\n"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Grade the outputs the same way for both models, then compute three numbers per model: pass rate, cost per passing task (not per request), and the 95th-percentile latency.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Day 5: test the failure modes, not just the averages.&lt;/strong&gt; Run your prompt-injection and tool-misuse cases against the candidate. A model that is better at agentic coding is also better at following instructions hidden in a web page or a ticket. My &lt;a href="https://kunalvohra.com/blog/ai-agent-security" rel="noopener noreferrer"&gt;AI agent security guide&lt;/a&gt; has the checklist; the tool permissions belong in your code, not in the model.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Days 6 to 7: decide.&lt;/strong&gt; Switch only if cost per passing task drops meaningfully and the pass rate holds within your tolerance. If it is close, keep the candidate wired in as a fallback or for one workload.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I'd do as a CTO this week
&lt;/h2&gt;

&lt;p&gt;I would not migrate anything because of a benchmark table. I would run the eval above on the one workload where frontier-model spend hurts most, through Moonshot's API or a host we already trust, with scrubbed data. If K3.1 Pro passes at half the cost per passing task, I'd move that one workload and keep the current model as fallback.&lt;/p&gt;

&lt;p&gt;I would close the self-hosting conversation in the same meeting. A 1.56 TB model is not a startup infrastructure project. If data must stay inside your own boundary, a smaller open model that fits on one or two GPUs, or a dedicated endpoint from a host with the right contracts, is the realistic path.&lt;/p&gt;

&lt;p&gt;And I would treat the "removed hacking skills" claim as a welcome step, not a security control. Your agent's blast radius is set by the tools and credentials you give it, whichever model sits in the middle.&lt;/p&gt;

&lt;p&gt;If you don't have someone senior to own model choices like this, it rarely needs a full-time hire. A &lt;a href="https://kunalvohra.com/technical-advisor" rel="noopener noreferrer"&gt;technical advisor&lt;/a&gt; retainer can review the eval and the vendor terms; a &lt;a href="https://kunalvohra.com/fractional-cto" rel="noopener noreferrer"&gt;fractional CTO&lt;/a&gt; can own the AI roadmap, the provider layer and the cost tracking end to end.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Is Kimi K3.1 Pro open source?&lt;/strong&gt; It is open-weight: the full weights are downloadable from Hugging Face under the Kimi K3.1 Pro License. That licence allows commercial use and fine-tuning, with conditions for model-as-a-service businesses above $20 million in yearly revenue and attribution for products above 100 million monthly users or $20 million in monthly revenue. It is not an OSI-approved licence such as Apache 2.0 or MIT.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is Kimi K3.1 Pro better than Claude Opus 5.5?&lt;/strong&gt; On Moonshot's own benchmarks it is close on agentic coding (65.7 against 66.4 on Terminal-Bench 4.0, ahead on DeepSWE and SWE-Marathon) and behind on some hard reasoning sets. Independent indexes rank it below the top closed models. Test it on your own tasks before deciding.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How much does the Kimi K3.1 Pro API cost?&lt;/strong&gt; On Moonshot's platform, $2 per million input tokens, $0.20 per million cached input tokens and $10 per million output tokens. The Highspeed tier is $4 input and $20 output.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Can I run Kimi K3.1 Pro locally?&lt;/strong&gt; Not on a laptop or a single server. The weights are 1.56 TB, and estimates for serving it start above 1.56 TB of combined GPU memory, which means multiple eight-GPU nodes. For local experiments, smaller open models are the realistic choice; my &lt;a href="https://kunalvohra.com/blog/how-to-run-airllm-on-mac" rel="noopener noreferrer"&gt;AirLLM on Mac guide&lt;/a&gt; shows what a Mac can and cannot do.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What hardware do I need to self-host Kimi K3.1 Pro?&lt;/strong&gt; Moonshot has not published an official configuration. Hosting providers estimate more than 1.56 TB to 2 TB or more of combined GPU memory, so roughly 20 or more 80 GB GPUs before headroom. The card names vLLM, SGLang and TokenSpeed as supported engines.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is it safe to send customer data to Kimi K3.1 Pro?&lt;/strong&gt; That depends on the provider's data terms and your contracts, not on the model. Read the retention and training terms of Moonshot's API or of the host you choose, check your sub-processor commitments, and test with scrubbed data first.&lt;/p&gt;

&lt;p&gt;If you're deciding whether to move part of your AI spend to a new model, &lt;a href="https://calendly.com/kunalvohra" rel="noopener noreferrer"&gt;book a 30-minute call&lt;/a&gt;. Bring your current cost per task and the workload that hurts most, and I'll help you design the eval and read the results.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://www.kimi.ai/blog/kimi-k3-1-pro" rel="noopener noreferrer"&gt;Kimi K3.1 Pro technical blog&lt;/a&gt; (Moonshot AI)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://huggingface.co/moonshotai/Kimi-K3.1-Pro" rel="noopener noreferrer"&gt;Kimi K3.1 Pro model card&lt;/a&gt; (Hugging Face, Moonshot AI)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://huggingface.co/moonshotai/Kimi-K3.1-Pro/blob/main/LICENSE" rel="noopener noreferrer"&gt;Kimi K3.1 Pro License&lt;/a&gt; (Moonshot AI)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://platform.kimi.ai/" rel="noopener noreferrer"&gt;Kimi API pricing&lt;/a&gt; (Moonshot AI)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://artificialanalysis.ai/models/kimi-k3-1-pro" rel="noopener noreferrer"&gt;Kimi K3.1 Pro intelligence, performance and price analysis&lt;/a&gt; (Artificial Analysis)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://thenextweb.com/news/moonshot-kimi-k3-1-pro-open-weight-cyber" rel="noopener noreferrer"&gt;Moonshot releases Kimi K3.1 Pro and says it cut its hacking skills&lt;/a&gt; (The Next Web)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://help.apiyi.com/en/kimi-k3-1-pro-open-weights-license-self-hosting-en.html" rel="noopener noreferrer"&gt;Kimi K3.1 Pro open weights: licence conditions and self-deployment threshold&lt;/a&gt; (APIYI)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.yottalabs.ai/post/kimi-k3-1-pro-specs-price-open-weights-benchmarks" rel="noopener noreferrer"&gt;Kimi K3.1 Pro specs, price, open weights date and benchmarks&lt;/a&gt; (Yotta Labs)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.kucoin.com/news/flash/moonshot-ai-s-kimi-k3-1-pro-opens-on-hugging-face-tops-open-weight-model-rankings" rel="noopener noreferrer"&gt;Kimi K3.1 Pro opens on Hugging Face&lt;/a&gt; (ChainCatcher via KuCoin)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.acronis.com/en/tru/posts/bossai-inside-the-ai-attack-mesh-that-targeted-taiwans-emergency-services/" rel="noopener noreferrer"&gt;BOSSAI: inside the AI attack mesh that targeted Taiwan's emergency services&lt;/a&gt; (Acronis)&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>programming</category>
      <category>machinelearning</category>
      <category>startup</category>
    </item>
    <item>
      <title>APT Vulnerability CVE-2026-109518 (November 2026): How to Fix Debian and Ubuntu Servers, Step by Step</title>
      <dc:creator>Kunal Vohra</dc:creator>
      <pubDate>Thu, 05 Nov 2026 12:00:00 +0000</pubDate>
      <link>https://dev.to/kunalvohra/apt-vulnerability-cve-2026-109518-november-2026-how-to-fix-debian-and-ubuntu-servers-step-by-2o81</link>
      <guid>https://dev.to/kunalvohra/apt-vulnerability-cve-2026-109518-november-2026-how-to-fix-debian-and-ubuntu-servers-step-by-2o81</guid>
      <description>&lt;p&gt;CVE-2026-109518 is a flaw in APT, the package manager on Debian, Ubuntu and their derivatives, that lets someone who can change the HTTP traffic between APT and the package server replace a package after APT has already checked its signature. Because APT installs packages as root, that means code running as root on your machine. It affects default setups from Debian 12 and Ubuntu 22.04 onwards. The fix is an updated &lt;code&gt;apt&lt;/code&gt; package, published on 4 November 2026 for Debian (2.6.2 for bookworm, 3.0.4 for trixie) and Ubuntu (22.04, 24.04, 26.04 and 26.10). Upgrade &lt;code&gt;apt&lt;/code&gt; itself first, on a network you trust, and then rebuild your Docker images.&lt;/p&gt;

&lt;p&gt;I'm Kunal Vohra, a fractional CTO and a security researcher and cybersecurity advisor. Yesterday I wrote about patching the &lt;a href="https://kunalvohra.com/blog/linux-kernel-cve-2026-98432-how-to-fix" rel="noopener noreferrer"&gt;Linux kernel TCP flaw CVE-2026-98432&lt;/a&gt;, and the first command in that guide was &lt;code&gt;sudo apt update&lt;/code&gt;. This bug is in the tool you use to apply every other fix, so it jumps the queue. Here is what it is, how to check where you stand, and how to patch without tripping over the bug while you do it.&lt;/p&gt;

&lt;h2&gt;
  
  
  What CVE-2026-109518 is
&lt;/h2&gt;

&lt;p&gt;CVE-2026-109518 is a "download verification bypass" in APT, which is how Debian's own changelog for the fix labels it. In plain terms: APT downloads a package, verifies it, and then installs it, and this bug opens a short window in which the file on disk can be changed between the check and the install.&lt;/p&gt;

&lt;p&gt;According to the reporting on Debian's advisory, the attacker uses HTTP redirects and ordinary temporary server errors to make APT retry a download. With the right redirects, APT ends up treating two different downloads as the same file, and the way it reuses files on disk when a download resumes lets the attacker's content replace the verified package. Debian credits the researcher Max Justicz with finding it. The upstream fix, in APT 3.3.6 by Julian Andres Klode, changes APT's download methods to "open target files exclusively", so two downloads can no longer end up writing to the same file.&lt;/p&gt;

&lt;p&gt;The facts that matter for prioritising it:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The attacker has to be on the path.&lt;/strong&gt; Someone who can modify the HTTP traffic between your machine and the package server: a malicious or compromised mirror, a proxy, a hostile network (public Wi-Fi, a compromised router), or anyone who can intercept traffic in between. A random attacker on the internet cannot trigger it against your server by itself.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It fires when you install or update.&lt;/strong&gt; Nothing happens while a server sits idle. The risk is every &lt;code&gt;apt update&lt;/code&gt;, &lt;code&gt;apt upgrade&lt;/code&gt; and &lt;code&gt;apt-get install&lt;/code&gt;, which includes every Docker build and CI job that installs packages.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Retries are the switch.&lt;/strong&gt; The bug needs APT's &lt;code&gt;Acquire::Retries&lt;/code&gt; setting to be above zero. It is 3 by default on Debian 12 and Ubuntu 22.04 LTS and later, which is why those releases are affected out of the box. Debian 11 and Ubuntu 20.04 are not exposed by default, but they are if someone set retries in a config file or script.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Severity: high, with no single score yet.&lt;/strong&gt; Ubuntu rates it high priority (CVSS 4.0 score of 7.7, as reported by The Hacker News); APT's developers call it critical; Debian's tracker lists it as important for testing. I could not find an NVD score at the time of writing.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No reported attacks, but a public proof of concept is indexed.&lt;/strong&gt; None of the advisories describe exploitation in the wild, and it is not in CISA's KEV catalogue. A proof-of-concept repository is listed in public PoC trackers, so treat this as "assume someone will try it".&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The line I'd put in front of any founder: &lt;strong&gt;the signature check is only as good as the moment you check it, and this bug moves the swap to just after that moment.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Is this you? A two-minute check
&lt;/h2&gt;

&lt;p&gt;You are affected if you run Debian 12 or 13, Ubuntu 22.04 or later, or a derivative (Linux Mint, Pop!_OS, Raspberry Pi OS, many cloud images), and you have not installed the fixed &lt;code&gt;apt&lt;/code&gt; yet. That covers a lot of startup infrastructure: web and API servers, bastion hosts, CI runners, Kubernetes nodes, and every container image built &lt;code&gt;FROM debian&lt;/code&gt; or &lt;code&gt;FROM ubuntu&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;On each machine, check which release and which &lt;code&gt;apt&lt;/code&gt; you have:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;cat /etc/os-release
dpkg-query -W -f=&lt;span class="s1"&gt;'${Version}\n'&lt;/span&gt; apt

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then check the two conditions the attack depends on, retries and plain HTTP sources:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;apt-config dump &lt;span class="o"&gt;|&lt;/span&gt; grep -i retries
grep -rhE &lt;span class="s1"&gt;'^(deb|URIs:)'&lt;/span&gt; /etc/apt/sources.list /etc/apt/sources.list.d/ 2&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;/dev/null

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If retries shows a number above 0 (or nothing, on Debian 12+ and Ubuntu 22.04+, where the default is 3) and your sources start with &lt;code&gt;http://&lt;/code&gt;, you have both preconditions. Many stock images still point at &lt;code&gt;http://&lt;/code&gt; mirror URLs, because APT relies on signatures rather than the transport for integrity. That assumption is exactly what this bug breaks.&lt;/p&gt;

&lt;p&gt;Now compare your &lt;code&gt;apt&lt;/code&gt; version with the fixed versions below.&lt;/p&gt;

&lt;h2&gt;
  
  
  Fixed apt versions for CVE-2026-109518
&lt;/h2&gt;

&lt;p&gt;These are the versions published on 4 November 2026. Debian's are confirmed on the Debian package tracker; Ubuntu's are as listed by Ubuntu and reported by The Hacker News, so check Ubuntu's CVE page for your release before you rely on them.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Debian&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Debian 13 (trixie, stable):&lt;/strong&gt; apt &lt;strong&gt;3.0.4&lt;/strong&gt; (DSA-6607-1). Vulnerable: 3.0.3.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Debian 12 (bookworm, oldstable):&lt;/strong&gt; apt &lt;strong&gt;2.6.2&lt;/strong&gt; (DLA-4880-1). Vulnerable: 2.6.1.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Debian testing (forky) and unstable (sid):&lt;/strong&gt; fixed upstream in apt &lt;strong&gt;3.3.6&lt;/strong&gt; , accepted into unstable on 4 November. Testing was still on vulnerable 3.3.5 when I checked, with the migration blocked.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Debian 11 (bullseye):&lt;/strong&gt; not exposed in its default configuration; The Hacker News reports a fix in 2.2.5 for systems where retries were turned on.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;Ubuntu&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Ubuntu 26.10:&lt;/strong&gt; apt &lt;strong&gt;3.3.3ubuntu0.1&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ubuntu 26.04 LTS:&lt;/strong&gt; apt &lt;strong&gt;3.2.1ubuntu0.1&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ubuntu 24.04 LTS:&lt;/strong&gt; apt &lt;strong&gt;2.8.4ubuntu0.1&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ubuntu 22.04 LTS:&lt;/strong&gt; apt &lt;strong&gt;2.4.15ubuntu0.1&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ubuntu 20.04 and older:&lt;/strong&gt; not exposed by default; fixes for non-default setups go through Ubuntu Pro (ESM). Ubuntu tracks eight releases in total for this CVE.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;One timing note: on the afternoon of 4 November, Ubuntu's tracker still marked releases as vulnerable for a while after the fixed builds were published. If your version matches the list above, you have the fix, whatever a cached page says.&lt;/p&gt;

&lt;h2&gt;
  
  
  How to fix CVE-2026-109518, step by step
&lt;/h2&gt;

&lt;p&gt;The awkward part: the tool you are patching is the tool doing the downloading. So the order matters. Turn off the precondition, upgrade &lt;code&gt;apt&lt;/code&gt; on its own, verify, and only then run your normal upgrades.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 1: Do the update from a network you trust
&lt;/h3&gt;

&lt;p&gt;Do not patch a laptop on hotel Wi-Fi, and don't route this update through a proxy or mirror you don't control. For cloud servers talking to your provider's or your distribution's official mirror, the practical risk during the update is low, but it is not zero, which is why Step 2 exists.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 2: Remove the retry precondition for the upgrade itself
&lt;/h3&gt;

&lt;p&gt;The bug needs &lt;code&gt;Acquire::Retries&lt;/code&gt; above zero, so run the &lt;code&gt;apt&lt;/code&gt; upgrade with retries set to 0 for that one command:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;sudo apt-get -o Acquire::Retries=0 update
sudo apt-get -o Acquire::Retries=0 install --only-upgrade apt

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If the download fails on a flaky connection, just run it again. A failed download is the price of not retrying automatically.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 3: Verify the fixed version
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;dpkg-query -W -f=&lt;span class="s1"&gt;'${Version}\n'&lt;/span&gt; apt

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Compare with the list above. No reboot is needed: &lt;code&gt;apt&lt;/code&gt; is a userspace tool, and the next run uses the new code.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 4: Now run your normal updates
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;sudo apt update
sudo apt full-upgrade

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If you were also waiting on the kernel fix from earlier this week, this is when you install it and reboot. The &lt;a href="https://kunalvohra.com/blog/linux-kernel-cve-2026-98432-how-to-fix" rel="noopener noreferrer"&gt;CVE-2026-98432 guide&lt;/a&gt; has the per-distribution details.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 5 (optional stopgap): disable retries on machines you can't patch today
&lt;/h3&gt;

&lt;p&gt;For a machine you cannot upgrade right away, the reporting on the advisories (not the vendors themselves) points to setting retries to 0 as a way to remove one required condition. It makes updates less forgiving on bad connections, so treat it as temporary:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;echo&lt;/span&gt; &lt;span class="s1"&gt;'Acquire::Retries "0";'&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt; sudo tee /etc/apt/apt.conf.d/99-cve-2026-109518

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Delete that file once the fixed &lt;code&gt;apt&lt;/code&gt; is installed if you want automatic retries back. Neither Debian nor Ubuntu offers an official workaround; the update is the fix.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 6: Prefer HTTPS mirrors where you can
&lt;/h3&gt;

&lt;p&gt;Modern APT supports &lt;code&gt;https://&lt;/code&gt; sources out of the box. If your mirror offers HTTPS, switching makes the interception this bug depends on much harder. To be precise about it: none of the advisories says HTTPS sources are immune, so this is defence in depth on top of the update, not instead of it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Docker images and CI runners: where the exposure really lives
&lt;/h2&gt;

&lt;p&gt;Production servers run &lt;code&gt;apt&lt;/code&gt; occasionally. Your build pipeline runs it all the time. Every &lt;code&gt;RUN apt-get update &amp;amp;&amp;amp; apt-get install&lt;/code&gt; in a Dockerfile, and every CI job that installs a package, is an APT session that this bug applies to, and it often runs on shared networks you have no visibility into.&lt;/p&gt;

&lt;p&gt;What I would do this week:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Find every image that installs packages.&lt;/strong&gt; Search your repositories for &lt;code&gt;apt-get&lt;/code&gt; and &lt;code&gt;apt install&lt;/code&gt; in Dockerfiles and CI config files.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Pull fresh base images and check what is inside them.&lt;/strong&gt; A base image tag only contains the fixed &lt;code&gt;apt&lt;/code&gt; once it has been rebuilt upstream. Check before you trust it:
&lt;/li&gt;
&lt;/ol&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;docker pull debian:trixie
docker run --rm debian:trixie dpkg-query -W -f=&lt;span class="s1"&gt;'${Version}\n'&lt;/span&gt; apt

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Until your base image ships the fix, upgrade apt first in the Dockerfile,&lt;/strong&gt; with retries off, before installing anything else:
&lt;/li&gt;
&lt;/ol&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="k"&gt;RUN&lt;/span&gt; apt-get -o Acquire::Retries=0 update \
 &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; apt-get -o Acquire::Retries=0 install -y --only-upgrade apt \
 &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; apt-get update &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; apt-get install -y --no-install-recommends your-packages \
 &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; rm -rf /var/lib/apt/lists/*

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Rebuild and redeploy&lt;/strong&gt; images built in the last few weeks, so the fix is in what actually runs, not just in your Dockerfile.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Patch self-hosted CI runners&lt;/strong&gt; the same way as servers, and check that your hosted CI provider has updated its runner images.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  What this means for startups
&lt;/h2&gt;

&lt;p&gt;This is a supply chain bug in the most trusted part of the stack. Most teams treat the package manager as plumbing: it has signatures, so it must be safe. CVE-2026-109518 is a reminder that "signed" protects the file at the moment it is checked, and that the boring parts of your pipeline (mirrors, proxies, CI networks, base images) are part of your attack surface.&lt;/p&gt;

&lt;p&gt;The teams that close this in a day are the ones with the same three things I mentioned for the kernel bug: an inventory of every Linux host and image, a patch routine that isn't a heroic effort, and someone who owns it. The last three big fixes on this blog, the &lt;a href="https://kunalvohra.com/blog/nextjs-cve-2026-96889-how-to-fix" rel="noopener noreferrer"&gt;Next.js CVE-2026-96889 upgrade&lt;/a&gt;, &lt;a href="https://kunalvohra.com/blog/php2shell-cve-2026-109618-how-to-fix" rel="noopener noreferrer"&gt;PHP2Shell&lt;/a&gt; and the kernel patch, all came down to the same lesson: the fix was quick, finding every place that needed it was the slow part. My &lt;a href="https://kunalvohra.com/blog/startup-security-checklist" rel="noopener noreferrer"&gt;startup security checklist&lt;/a&gt; puts that inventory near the top for exactly that reason.&lt;/p&gt;

&lt;p&gt;If nobody on your team owns this kind of work, that is the real gap. A &lt;a href="https://kunalvohra.com/technical-advisor" rel="noopener noreferrer"&gt;technical advisor&lt;/a&gt; can set up the inventory, base-image policy and patch routine with your engineers in a few sessions; a &lt;a href="https://kunalvohra.com/fractional-cto" rel="noopener noreferrer"&gt;fractional CTO&lt;/a&gt; owns it on an ongoing basis, along with the rest of your security posture.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;What is CVE-2026-109518?&lt;/strong&gt; A download verification bypass in APT, the package manager on Debian and Ubuntu. An attacker who can modify the HTTP traffic between APT and the package server can use redirects and retries to swap a package after APT has verified it, which leads to code running as root when the package installs. It was disclosed and patched on 4 November 2026.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is Ubuntu affected by CVE-2026-109518?&lt;/strong&gt; Yes. Ubuntu 22.04 LTS and later are affected in their default configuration because APT retries downloads by default. Fixed apt builds are 2.4.15ubuntu0.1 (22.04), 2.8.4ubuntu0.1 (24.04), 3.2.1ubuntu0.1 (26.04) and 3.3.3ubuntu0.1 (26.10). Ubuntu 20.04 is not exposed by default.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How do I fix the APT vulnerability on Debian?&lt;/strong&gt; Upgrade the apt package: 3.0.4 on Debian 13 (trixie), 2.6.2 on Debian 12 (bookworm). Run &lt;code&gt;sudo apt-get -o Acquire::Retries=0 update&lt;/code&gt; and &lt;code&gt;sudo apt-get -o Acquire::Retries=0 install --only-upgrade apt&lt;/code&gt; from a trusted network, then confirm with &lt;code&gt;dpkg-query -W apt&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is CVE-2026-109518 being exploited?&lt;/strong&gt; No attacks had been reported by Debian, Ubuntu or CISA as of 4 to 5 November 2026. A proof-of-concept repository is listed in public trackers, so patch now rather than waiting for reports.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Does using HTTPS mirrors protect me?&lt;/strong&gt; It makes the attack harder, because the attacker has to tamper with traffic in transit, but no advisory says HTTPS sources are immune. Use HTTPS where your mirror supports it and still install the fixed apt.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Do I need to reboot after updating apt?&lt;/strong&gt; No. APT is a userspace program, so the fixed version is used the next time it runs. You only need a reboot if you also install a new kernel.&lt;/p&gt;

&lt;p&gt;If you're not sure how many servers, runners and images in your stack run APT, or you want a patch routine that makes the next one of these boring, &lt;a href="https://calendly.com/kunalvohra" rel="noopener noreferrer"&gt;book a 30-minute call&lt;/a&gt;. Bring your Dockerfiles; that's usually where the surprises are.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://tracker.debian.org/pkg/apt" rel="noopener noreferrer"&gt;apt package news and versions&lt;/a&gt; (Debian Package Tracker)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://metadata.ftp-master.debian.org/changelogs//main/a/apt/unstable_changelog" rel="noopener noreferrer"&gt;apt unstable changelog, 3.3.6 entry&lt;/a&gt; (Debian)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://security-tracker.debian.org/tracker/CVE-2026-109518" rel="noopener noreferrer"&gt;CVE-2026-109518&lt;/a&gt; (Debian Security Tracker)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://ubuntu.com/security/CVE-2026-109518" rel="noopener noreferrer"&gt;CVE-2026-109518&lt;/a&gt; (Ubuntu Security)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://thehackernews.com/2026/11/debian-and-ubuntu-apt-flaw-lets-mitm.html" rel="noopener noreferrer"&gt;Debian and Ubuntu APT Flaw Lets MitM Attackers Swap Verified Packages and Run Code as Root&lt;/a&gt; (The Hacker News)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://cybersecuritynews.com/debian-apt-flaw-lets-hackers-hijack-updates/" rel="noopener noreferrer"&gt;Debian and Ubuntu APT Flaw Lets Hackers Hijack Software Updates and Execute Code as Root&lt;/a&gt; (Cyber Security News)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://linuxiac.com/apt-3-3-6-released-with-security-fix-for-download-verification-bypass/" rel="noopener noreferrer"&gt;APT 3.3.6 Released with Security Fix for Download Verification Bypass&lt;/a&gt; (Linuxiac)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://runtimewire.com/article/debian-apt-cve-2026-109518-max-justicz" rel="noopener noreferrer"&gt;Debian patches APT flaw that let network attackers swap verified packages for root code&lt;/a&gt; (RuntimeWire)&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>security</category>
      <category>cybersecurity</category>
      <category>webdev</category>
      <category>startup</category>
    </item>
    <item>
      <title>Linux Kernel TCP Vulnerability CVE-2026-98432 (November 2026): What It Is and How to Patch Your Servers, Step by Step</title>
      <dc:creator>Kunal Vohra</dc:creator>
      <pubDate>Wed, 04 Nov 2026 12:00:00 +0000</pubDate>
      <link>https://dev.to/kunalvohra/linux-kernel-tcp-vulnerability-cve-2026-98432-november-2026-what-it-is-and-how-to-patch-your-53fg</link>
      <guid>https://dev.to/kunalvohra/linux-kernel-tcp-vulnerability-cve-2026-98432-november-2026-what-it-is-and-how-to-patch-your-53fg</guid>
      <description>&lt;p&gt;CVE-2026-98432 is a use-after-free in the Linux kernel's IPv4 TCP connection setup that a remote attacker can trigger without logging in, and that vendors say can lead to remote code execution. It has been in the kernel since Linux 4.4, so if you run Linux servers with any open IPv4 TCP port (a web server, an API, SSH), assume you are affected until you have checked your kernel. The fix is a kernel update and a reboot. Debian, Amazon Linux, Oracle Linux and RHEL have fixed kernels out; Ubuntu's fixes are still in its -proposed pocket as of 3 November 2026.&lt;/p&gt;

&lt;p&gt;I'm Kunal Vohra, a fractional CTO and a security researcher and cybersecurity advisor. Early-stage startups very often run on Ubuntu or Amazon Linux behind a cloud load balancer, with nobody whose job it is to reboot servers. This is the guide I'd hand that team: what the bug is, how to tell where you stand in five minutes, and how to patch each major distribution.&lt;/p&gt;

&lt;h2&gt;
  
  
  What CVE-2026-98432 is
&lt;/h2&gt;

&lt;p&gt;CVE-2026-98432 is a race condition in how the kernel hands IP options from an incoming TCP connection request to the new socket. When a server accepts an IPv4 TCP connection, the kernel passes the request's IP options to the child socket without copying them. If the client retransmits its connection request (a second SYN) at the wrong moment, the kernel can read those options after they have been freed. The upstream fix, merged on 5 October 2026, delays freeing them until it is safe.&lt;/p&gt;

&lt;p&gt;The facts that matter for prioritising it:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Remote and unauthenticated.&lt;/strong&gt; The trigger is a TCP connection request carrying IPv4 Record Route or Timestamp options, sent to any listening TCP port. No credentials, no user interaction.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;IPv4 only on the main path.&lt;/strong&gt; IPv6 is not affected by the main issue; a related IPv6 migration issue is being handled separately.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Old.&lt;/strong&gt; Present since Linux 4.4, so almost every supported distribution kernel is in range.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Scored differently by different vendors.&lt;/strong&gt; The kernel CVE team and Ubuntu rate it 9.8 (Critical). Red Hat first rated it lower and now treats it as Critical. Oracle and SUSE use 8.1, and Amazon Linux's CVE page shows 8.1. The spread reflects how hard the race is to win, not doubt about whether it is real.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A working exploit exists, privately.&lt;/strong&gt; Anthropic, whose researcher is credited in the kernel fix, published research on 27 October 2026 describing an exploit its model wrote for one configuration (a server with a Multipath TCP listener): 48 minutes to leak a kernel address, about five more minutes to run commands as root. The exploit code was not released. As of 3 November, no vendor or CERT has reported attacks in the wild.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The line I'd put in front of any founder: &lt;strong&gt;a 53-minute path from "open port" to "root on your server" now exists on paper, and the only complete fix is a kernel update plus a reboot.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Is your server affected? A five-minute check
&lt;/h2&gt;

&lt;p&gt;If a server runs a Linux kernel from 4.4 onwards, accepts IPv4 TCP connections, and has not been rebooted into a fixed kernel, treat it as affected. Containers don't change this: they share the host kernel, so the host is what you patch. Managed platforms where you never touch a kernel (serverless functions, most PaaS) are your provider's job; check their status page, but you have nothing to install.&lt;/p&gt;

&lt;p&gt;On each server, start with the running kernel and your distribution:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;uname&lt;/span&gt; &lt;span class="nt"&gt;-r&lt;/span&gt;
&lt;span class="nb"&gt;cat&lt;/span&gt; /etc/os-release

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then compare against your distribution's fixed version below. Two traps catch people here:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Installed is not running.&lt;/strong&gt; A fixed kernel package does nothing until you reboot into it. &lt;code&gt;uname -r&lt;/code&gt; shows what is actually running.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The MPTCP check is not an exposure test.&lt;/strong&gt; Red Hat mentions &lt;code&gt;ss -Mlnp&lt;/code&gt; to find Multipath TCP listeners, which matches the one configuration Anthropic exploited. No MPTCP listeners does not mean you are safe; the bug is in the ordinary IPv4 TCP path.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Make a list of every Linux host you own: web and API servers, databases, bastion hosts, CI runners, Kubernetes nodes. The last two are the ones teams forget.&lt;/p&gt;

&lt;h2&gt;
  
  
  How to fix CVE-2026-98432, step by step
&lt;/h2&gt;

&lt;p&gt;The fix is the same everywhere: install your distribution's patched kernel, reboot, and confirm the running version. What differs is whether a fixed kernel exists yet. Status below is as published by each vendor between 1 and 3 November 2026; check the vendor pages in the sources before you act, because they are moving daily.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 1: Debian
&lt;/h3&gt;

&lt;p&gt;Debian has fixed kernels in its security repositories: bookworm (Debian 12) in &lt;code&gt;linux&lt;/code&gt; 6.1.190-1 (and &lt;code&gt;linux-6.12&lt;/code&gt; 6.12.113-1~deb12u1), trixie (Debian 13) in 6.12.113-1 or later.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;sudo &lt;/span&gt;apt update
&lt;span class="nb"&gt;sudo &lt;/span&gt;apt full-upgrade
&lt;span class="nb"&gt;sudo &lt;/span&gt;reboot
&lt;span class="c"&gt;# after the reboot&lt;/span&gt;
&lt;span class="nb"&gt;uname&lt;/span&gt; &lt;span class="nt"&gt;-r&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Step 2: Amazon Linux
&lt;/h3&gt;

&lt;p&gt;Amazon Linux shows the CVE as fixed on every kernel it ships, with advisories dated 24 to 30 October 2026 (ALAS2023-2026-3365 for the Amazon Linux 2023 &lt;code&gt;kernel&lt;/code&gt; package, ALAS2-2026-4065 for Amazon Linux 2 core, plus separate advisories for the 5.4, 5.10 and 5.15 extras and the AL2023 kernel6.12 and kernel6.18 packages).&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# Amazon Linux 2023&lt;/span&gt;
&lt;span class="nb"&gt;sudo &lt;/span&gt;dnf upgrade &lt;span class="nt"&gt;--refresh&lt;/span&gt; kernel
&lt;span class="c"&gt;# Amazon Linux 2&lt;/span&gt;
&lt;span class="nb"&gt;sudo &lt;/span&gt;yum update kernel
&lt;span class="nb"&gt;sudo &lt;/span&gt;reboot
&lt;span class="nb"&gt;uname&lt;/span&gt; &lt;span class="nt"&gt;-r&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If you run AWS-managed node groups or launch templates, rolling to the latest AMI is often cleaner than patching nodes in place.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 3: RHEL, Rocky, Alma and Oracle Linux
&lt;/h3&gt;

&lt;p&gt;Red Hat shipped fixed kernels on 1 November 2026: RHEL 8.10 in &lt;code&gt;kernel-4.18.0-553.182.1.el8_10&lt;/code&gt; (RHSA-2026:93414), RHEL 9.8 in &lt;code&gt;kernel-5.14.0-687.64.1.el9_8&lt;/code&gt; (RHSA-2026:93412) and RHEL 10.2 in &lt;code&gt;kernel-6.12.0-211.72.1.el10_2&lt;/code&gt; (RHSA-2026:93413). Red Hat's errata say the system must be rebooted. RHEL 7 is not affected. Oracle Linux's UEK kernels were fixed on 16 October. Rebuilds such as Rocky and Alma follow Red Hat; check their own advisories for the matching build.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;sudo &lt;/span&gt;dnf upgrade &lt;span class="nt"&gt;--refresh&lt;/span&gt; kernel
&lt;span class="nb"&gt;sudo &lt;/span&gt;reboot
&lt;span class="nb"&gt;uname&lt;/span&gt; &lt;span class="nt"&gt;-r&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;On OpenShift, node kernels come from RHEL CoreOS; follow Red Hat's OpenShift guidance in the bulletin rather than patching nodes by hand.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 4: Ubuntu
&lt;/h3&gt;

&lt;p&gt;Ubuntu is the one most startups run, and the one without a fix in the normal update channel yet. As of Canonical's 2 November post, fixed kernels for 22.04 LTS (5.15.0-203.213), 24.04 LTS (6.8.0-149.149) and 26.04 LTS (7.0.0-43.43) are in the -proposed pocket while Canonical runs qualification testing; 26.10 has its fix in 7.3.0-10.10. Fixes for 16.04, 18.04 and 20.04 are pending and need Ubuntu Pro (ESM). Cloud, FIPS and real-time kernel flavours are still in progress, so check the CVE page for linux-aws, linux-azure or linux-gcp if that is what &lt;code&gt;uname -r&lt;/code&gt; shows.&lt;/p&gt;

&lt;p&gt;You have two choices. The conservative one: apply the mitigation in Step 5 now, then run &lt;code&gt;sudo apt update &amp;amp;&amp;amp; sudo apt full-upgrade&lt;/code&gt; and reboot the day the fix reaches the normal updates pocket. The faster one, which Canonical documents: install the kernel from -proposed explicitly.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;uname&lt;/span&gt; &lt;span class="nt"&gt;-r&lt;/span&gt;
dpkg &lt;span class="nt"&gt;-l&lt;/span&gt; &lt;span class="s1"&gt;'linux-image*'&lt;/span&gt; | &lt;span class="nb"&gt;grep&lt;/span&gt; ^ii

&lt;span class="nb"&gt;sudo &lt;/span&gt;add-apt-repository &lt;span class="nt"&gt;--component&lt;/span&gt; main &lt;span class="nt"&gt;--pocket&lt;/span&gt; proposed http://archive.ubuntu.com/ubuntu
&lt;span class="nb"&gt;sudo &lt;/span&gt;apt update
dpkg-query &lt;span class="nt"&gt;-W&lt;/span&gt; &lt;span class="nt"&gt;-f&lt;/span&gt; &lt;span class="s1"&gt;'${source:Package}\t${binary:Package}\n'&lt;/span&gt; | &lt;span class="nb"&gt;awk&lt;/span&gt; &lt;span class="s1"&gt;'$1 ~ "^linux-meta" { print $2 }'&lt;/span&gt; | xargs &lt;span class="nb"&gt;sudo &lt;/span&gt;apt &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;-t&lt;/span&gt; &lt;span class="si"&gt;$(&lt;/span&gt;&lt;span class="nb"&gt;.&lt;/span&gt; /etc/os-release &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="nb"&gt;echo&lt;/span&gt; &lt;span class="nv"&gt;$VERSION_CODENAME&lt;/span&gt;&lt;span class="si"&gt;)&lt;/span&gt;&lt;span class="nt"&gt;-proposed&lt;/span&gt; &lt;span class="nt"&gt;--only-upgrade&lt;/span&gt;
&lt;span class="nb"&gt;sudo &lt;/span&gt;reboot

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Packages in -proposed are still being tested, so I'd do this on one non-critical server first, watch it for a day, then roll it out. Packages from -proposed are not picked for upgrade by default, which is why the explicit install is needed.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 5: the nftables stopgap (only until you can patch)
&lt;/h3&gt;

&lt;p&gt;If you can't reboot into a fixed kernel today, Red Hat publishes a filter that drops inbound IPv4 TCP packets carrying IP options, which blocks the known trigger. Save this as &lt;code&gt;cve-2026-98432.nft&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;table inet cve_2026_98432 {
    chain input {
        type filter hook input priority -310; policy accept;
        ip protocol tcp ip hdrlength &amp;gt; 5 counter drop
    }
}

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then check, load and confirm it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;sudo &lt;/span&gt;nft &lt;span class="nt"&gt;--check&lt;/span&gt; &lt;span class="nt"&gt;--file&lt;/span&gt; cve-2026-98432.nft
&lt;span class="nb"&gt;sudo &lt;/span&gt;nft &lt;span class="nt"&gt;--file&lt;/span&gt; cve-2026-98432.nft
&lt;span class="nb"&gt;sudo &lt;/span&gt;nft list table inet cve_2026_98432

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Know its limits before you rely on it. Red Hat and Canonical both say filtering reduces exposure but is not a substitute for patching. It only protects the host it is loaded on: traffic between containers on the same host can bypass it, and on Kubernetes or OpenShift pod-to-pod traffic is not covered. It can break the rare workloads that depend on IPv4 options, so test it outside production first. It does not survive a reboot unless you add it to your nftables configuration. When the fixed kernel is running, remove it:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;sudo &lt;/span&gt;nft delete table inet cve_2026_98432

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;One mitigation I'd skip: disabling Multipath TCP. Red Hat and Canonical both say the bug is not specific to MPTCP, and on RHEL MPTCP is already off by default, so turning it off changes nothing on a default system.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 6: confirm, and don't forget the long tail
&lt;/h3&gt;

&lt;p&gt;After each reboot, &lt;code&gt;uname -r&lt;/code&gt; should show the fixed version. Then work down the list you made earlier: CI runners (self-hosted ones especially), bastion hosts, database servers you manage yourself, and Kubernetes nodes. Custom kernels need 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.56, 7.2.10, or 7.3 or later; the kernel team's advice is to move to the latest stable release rather than cherry-pick.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this means for startups
&lt;/h2&gt;

&lt;p&gt;This bug is a reminder that "we're on managed cloud" is not the same as "someone patches our servers." A cloud VM is your operating system to maintain. The teams that will be done with CVE-2026-98432 by Friday are the ones that already had three boring things in place:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;An inventory.&lt;/strong&gt; A list of every Linux host and what kernel it runs. Without it, you find the forgotten server last, or never.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Reboots that aren't scary.&lt;/strong&gt; Two instances behind a load balancer, or nodes you can drain and replace, so a kernel update is a routine rolling restart rather than a downtime window someone has to approve.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Automatic security updates where you can tolerate them,&lt;/strong&gt; with &lt;code&gt;unattended-upgrades&lt;/code&gt; on Ubuntu and Debian or &lt;code&gt;dnf-automatic&lt;/code&gt; on RHEL-family systems, plus a scheduled reboot window. Updates that install but never reboot leave you exactly as exposed as before.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;There is a second lesson, and it is the reason this bug got attention. An AI model wrote a working root exploit for a race condition in under an hour. Whatever you think of that, the gap between "patch published" and "exploit exists" is shrinking, and the planning assumption that you have weeks to get round to a kernel update is getting weaker. My &lt;a href="https://kunalvohra.com/blog/startup-security-checklist" rel="noopener noreferrer"&gt;startup security checklist&lt;/a&gt; puts patching and tested backups near the top for that reason, and the same pattern showed up last month with the &lt;a href="https://kunalvohra.com/blog/nextjs-cve-2026-96889-how-to-fix" rel="noopener noreferrer"&gt;Next.js CVE-2026-96889 fix&lt;/a&gt; and &lt;a href="https://kunalvohra.com/blog/php2shell-cve-2026-109618-how-to-fix" rel="noopener noreferrer"&gt;PHP2Shell&lt;/a&gt;: the fix was quick, finding every place that needed it was the slow part.&lt;/p&gt;

&lt;p&gt;If nobody on your team owns this kind of work, that is the gap to close, not this one CVE. A &lt;a href="https://kunalvohra.com/technical-advisor" rel="noopener noreferrer"&gt;technical advisor&lt;/a&gt; can set up the inventory and the patch routine with your engineers in a few sessions; a &lt;a href="https://kunalvohra.com/fractional-cto" rel="noopener noreferrer"&gt;fractional CTO&lt;/a&gt; owns it on an ongoing basis along with the rest of your security posture.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;What is CVE-2026-98432?&lt;/strong&gt; A use-after-free race in the Linux kernel's IPv4 TCP connection setup, present since Linux 4.4. A remote, unauthenticated attacker sending connection requests with IPv4 options to an open TCP port can trigger it, and vendors say it can lead to remote code execution. The fix is a kernel update and a reboot.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is Ubuntu affected by CVE-2026-98432, and is there a fix?&lt;/strong&gt; Yes. As of 2 to 3 November 2026, fixed kernels for Ubuntu 22.04, 24.04 and 26.04 LTS are in the -proposed pocket while Canonical tests them, with no date yet for the normal updates channel; 16.04, 18.04 and 20.04 fixes are pending. Use the nftables stopgap or install from -proposed carefully until the fix reaches updates.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is this vulnerability being exploited?&lt;/strong&gt; No attacks had been reported by vendors or CERTs as of 3 November 2026. Anthropic has described a private working exploit for one configuration, and proof-of-concept code that triggers the race (without code execution) is public, so I would not wait for reports of attacks before patching.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Does disabling MPTCP fix CVE-2026-98432?&lt;/strong&gt; No. The bug is in the ordinary IPv4 TCP path; MPTCP is only the configuration one exploit used. Red Hat and Canonical both say disabling it is not a complete mitigation.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Do I need to reboot after the kernel update?&lt;/strong&gt; Yes. The fixed kernel only protects you once it is running. Check with &lt;code&gt;uname -r&lt;/code&gt; after the reboot.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Are Docker containers and Kubernetes pods affected?&lt;/strong&gt; Containers share the host's kernel, so patch the host or the node. Host-level nftables rules don't cover traffic between containers or pods on the same host, which is another reason the kernel update is the real fix.&lt;/p&gt;

&lt;p&gt;If you're not sure which of your servers are exposed, or you want a patching routine that makes the next one boring, &lt;a href="https://calendly.com/kunalvohra" rel="noopener noreferrer"&gt;book a 30-minute call&lt;/a&gt;. Bring your list of hosts, or the fact that you don't have one, and we'll start there.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://access.redhat.com/security/vulnerabilities/RHSB-2026-016" rel="noopener noreferrer"&gt;RHSB-2026-016: Use-after-free in kernel IPv4 (CVE-2026-98432)&lt;/a&gt; (Red Hat)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://access.redhat.com/errata/RHSA-2026:93412" rel="noopener noreferrer"&gt;RHSA-2026:93412 kernel security update for RHEL 9&lt;/a&gt; (Red Hat)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://ubuntu.com/security/CVE-2026-98432" rel="noopener noreferrer"&gt;CVE-2026-98432&lt;/a&gt; (Ubuntu Security)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://ubuntu.com/blog/fixes-available-in-proposed-pocket-for-cve-2026-98432" rel="noopener noreferrer"&gt;Fixes available in -proposed pocket for CVE-2026-98432&lt;/a&gt; (Canonical)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://security-tracker.debian.org/tracker/CVE-2026-98432" rel="noopener noreferrer"&gt;CVE-2026-98432&lt;/a&gt; (Debian Security Tracker)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://explore.alas.aws.amazon.com/CVE-2026-98432.html" rel="noopener noreferrer"&gt;CVE-2026-98432&lt;/a&gt; (Amazon Linux Security Center)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.anthropic.com/research/discovering-vulnerabilities-with-claude" rel="noopener noreferrer"&gt;Discovering and exploiting vulnerabilities in critical, real-world software with Claude&lt;/a&gt; (Anthropic)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://thehackernews.com/2026/11/red-hat-patches-98-rated-linux-tcp-rce.html" rel="noopener noreferrer"&gt;Red Hat Patches 9.8-Rated Linux TCP RCE Flaw While Ubuntu Releases Fixes for Testing&lt;/a&gt; (The Hacker News)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://thehackernews.com/2026/10/linux-kernel-tcp-flaw-with-98-cvss.html" rel="noopener noreferrer"&gt;Linux kernel TCP flaw with 9.8 CVSS&lt;/a&gt; (The Hacker News)&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>security</category>
      <category>cybersecurity</category>
      <category>webdev</category>
      <category>startup</category>
    </item>
    <item>
      <title>Best Tech Stack for an MVP in 2026: How to Choose One You Won't Have to Rebuild</title>
      <dc:creator>Kunal Vohra</dc:creator>
      <pubDate>Fri, 23 Oct 2026 12:00:00 +0000</pubDate>
      <link>https://dev.to/kunalvohra/best-tech-stack-for-an-mvp-in-2026-how-to-choose-one-you-wont-have-to-rebuild-5ehh</link>
      <guid>https://dev.to/kunalvohra/best-tech-stack-for-an-mvp-in-2026-how-to-choose-one-you-wont-have-to-rebuild-5ehh</guid>
      <description>&lt;p&gt;The best tech stack for an MVP is the most boring one that does the job: a mainstream language your first engineer already knows, one database, one deployable application, and managed services for everything that is not your product (auth, payments, email, hosting). In 2026 that usually means TypeScript on the web, a single backend in TypeScript or Python, Postgres, and a cross-platform mobile framework if you need an app at all. The framework matters far less than founders think; what matters is whether the next person can pick it up in days.&lt;/p&gt;

&lt;p&gt;I have been the technical co-founder six times and I run &lt;a href="https://kunalvohra.com/mvp-development" rel="noopener noreferrer"&gt;MVP builds&lt;/a&gt; with my Panicle Tech team. This is the version I give non-technical founders: what to pick, what to avoid, and how to tell whether someone else's stack recommendation is honest.&lt;/p&gt;

&lt;h2&gt;
  
  
  What makes the best tech stack for an MVP?
&lt;/h2&gt;

&lt;p&gt;The best MVP tech stack is the one that gets a real product in front of users fastest and can be handed to a new engineer without a rewrite. Speed of the first build and ease of the second person are the only two tests that matter at this stage.&lt;/p&gt;

&lt;p&gt;Everything else people argue about (raw performance, elegance, what is trending) is a problem you earn later by having users. The stack's job is to stay out of the way of the one question the MVP exists to answer.&lt;/p&gt;

&lt;p&gt;The line I keep coming back to: &lt;strong&gt;the right stack at pre-seed is the boring one you can hire for.&lt;/strong&gt; I have watched interesting technology choices make a first engineering hire twice as hard, and at this stage that is an expensive kind of fun. If your first hire spends six weeks learning your stack before shipping anything, you paid for a framework with runway.&lt;/p&gt;

&lt;h2&gt;
  
  
  The rule that matters more than the framework
&lt;/h2&gt;

&lt;p&gt;Choose for hiring, handover and managed services, in that order. A stack that scores well on those three will outlive a "better" stack that scores badly on them.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Hiring.&lt;/strong&gt; Pick languages and frameworks with a deep pool of engineers in the markets you will hire from. If you will hire in the US, India and Eastern Europe, the overlap is TypeScript, Python, React and Postgres. A niche language might be lovely; it also shrinks your candidate pool and raises every salary in it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Handover.&lt;/strong&gt; Your MVP will change hands, from an agency or freelancer to your first engineer or a &lt;a href="https://kunalvohra.com/fractional-cto" rel="noopener noreferrer"&gt;fractional CTO&lt;/a&gt;. A mainstream stack with a real README and documented setup means the next person is productive in days.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Managed services.&lt;/strong&gt; Every piece of infrastructure you run yourself is a piece someone has to patch at 2 a.m. Managed services cost a little more per month and save you the person you cannot afford yet. At MVP stage a typical stack runs $100 to $500 a month in services, which is far cheaper than an engineer's week.&lt;/p&gt;

&lt;h2&gt;
  
  
  The MVP tech stack I point most founders at in 2026
&lt;/h2&gt;

&lt;p&gt;For a typical web SaaS or marketplace MVP, this is the default I start from, then adjust for the product. None of these are exotic, and that is the point.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Frontend: TypeScript and React, usually with Next.js.&lt;/strong&gt; React has the largest hiring pool of any frontend framework, TypeScript catches a whole class of bugs before users do, and Next.js handles routing, server rendering and SEO pages without extra setup. (This site runs on it.)&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Backend: one application, in TypeScript or Python.&lt;/strong&gt; If the product is mostly forms, workflows and data, a TypeScript backend keeps one language across the codebase. If the product is AI-heavy or data-heavy, Python is the better fit because that is where the model and data tooling lives. Pick one. A frontend in one language plus three backend services in two more is a team-of-ten architecture.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Database: Postgres, managed.&lt;/strong&gt; It handles relational data, JSON, full-text search and, with the pgvector extension, embeddings for retrieval. That covers almost every MVP I scope. Start with one managed Postgres instance and stay there until it genuinely hurts.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Auth, payments, email: buy them.&lt;/strong&gt; A managed auth provider, Stripe (or Razorpay for India-first products) for payments, and a transactional email service. Building your own login system burns budget on work no user notices, and it is a security liability on top.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Hosting: a managed platform.&lt;/strong&gt; A platform that deploys from your repository on every merge, with a managed database beside it. Kubernetes, multi-region failover and custom container orchestration are Series B problems.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Monitoring from day one.&lt;/strong&gt; Error tracking and uptime alerts in the first week, so you hear about breakage before your customers tell you.&lt;/p&gt;

&lt;h2&gt;
  
  
  Mobile: do you need a native app for your MVP?
&lt;/h2&gt;

&lt;p&gt;Usually not. If you need mobile at all, build one cross-platform app (React Native or Flutter) rather than separate iOS and Android codebases. If your users can live with a mobile-friendly web app first, start there and add the store apps once the idea is validated.&lt;/p&gt;

&lt;p&gt;Two native codebases means two builds, two sets of bugs and two release cycles for a product you have not validated yet. A single cross-platform build covers both stores from one codebase, and if your web frontend is already React, React Native keeps the team in one language. That is why my &lt;a href="https://kunalvohra.com/mvp-cost-calculator" rel="noopener noreferrer"&gt;MVP cost calculator&lt;/a&gt; treats mobile as one cross-platform build and adds app store deployment automatically.&lt;/p&gt;

&lt;p&gt;The exceptions are narrow: heavy use of device hardware, demanding graphics, or deep platform features that cross-platform frameworks reach late. For everyone else, native mobile before the web version has validated the idea is one of the first things I cut.&lt;/p&gt;

&lt;h2&gt;
  
  
  The best tech stack for an AI MVP
&lt;/h2&gt;

&lt;p&gt;An AI MVP uses the same boring foundation plus four specific choices: one model provider behind a thin seam, Postgres with pgvector before any dedicated vector database, cost logging on every model call, and a small eval script. The model is rarely what kills an AI MVP; the plumbing around it is.&lt;/p&gt;

&lt;p&gt;I run ColadAI, a multi-LLM platform, so these choices come from paying an inference bill every month.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;One provider, one seam.&lt;/strong&gt; Do not build a multi-provider routing layer into an MVP; that is premature. Do put a single thin interface between your product and the vendor, so swapping models later is a day rather than a month. My &lt;a href="https://kunalvohra.com/blog/gemini-vs-claude-vs-grok-vs-chatgpt-api-for-business" rel="noopener noreferrer"&gt;Gemini vs Claude vs Grok vs ChatGPT API guide&lt;/a&gt; covers choosing the first provider.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Use an API, not your own GPUs.&lt;/strong&gt; For almost every MVP the hosted API is cheaper and faster than self-hosting an open-weight model; the break-even math is in &lt;a href="https://kunalvohra.com/blog/self-hosting-llm-vs-api-cost" rel="noopener noreferrer"&gt;self-hosting an LLM vs using an API&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Retrieval only if you need it.&lt;/strong&gt; If the knowledge your product needs fits in a context window, you may not need retrieval at all. If you do, pgvector in the Postgres you already have is enough for an MVP. Fine-tuning almost never belongs in v1; &lt;a href="https://kunalvohra.com/blog/rag-vs-fine-tuning" rel="noopener noreferrer"&gt;RAG vs fine-tuning&lt;/a&gt; explains why.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Cost logging and evals from the first commit.&lt;/strong&gt; Log tokens in, tokens out, model and session for every call so you can answer what a user costs on launch day, and keep a golden set of twenty to fifty test cases with a script that scores them. Both are cheap to build in and painful to retrofit. The full cost picture is in &lt;a href="https://kunalvohra.com/blog/ai-mvp-cost" rel="noopener noreferrer"&gt;AI MVP cost&lt;/a&gt;, and if your product can call tools, read &lt;a href="https://kunalvohra.com/blog/ai-agent-security" rel="noopener noreferrer"&gt;AI agent security&lt;/a&gt; before you wire them up. I build MVPs for AI startups on exactly this shape; the &lt;a href="https://kunalvohra.com/mvp-development/ai-startups" rel="noopener noreferrer"&gt;AI startups page&lt;/a&gt; lists what goes in v1 and what waits.&lt;/p&gt;

&lt;h2&gt;
  
  
  No-code vs custom code for your MVP
&lt;/h2&gt;

&lt;p&gt;No-code and AI-assisted builders are a legitimate MVP stack in 2026 when the goal is to test demand, not to build the company on it. Choose custom code when the product's value is in logic the platform cannot express, when you handle sensitive data, or when you already know the MVP will become the real product.&lt;/p&gt;

&lt;p&gt;I build with AI tools daily, and they have moved the line: a working prototype in front of users can now take days. The costs arrive later: platform lock-in, per-user pricing that grows against you, limits on what you can integrate, and a rebuild the moment you need something the platform did not anticipate. Treat a no-code MVP as an experiment you expect to replace, and budget for the replacement.&lt;/p&gt;

&lt;p&gt;A practical test I use: if you would be embarrassed to show the codebase to an investor's technical reviewer, it is a prototype, not an MVP. That is fine, as long as you know which one you have. My guide to &lt;a href="https://kunalvohra.com/blog/how-to-build-an-mvp-without-a-technical-cofounder" rel="noopener noreferrer"&gt;building an MVP without a technical co-founder&lt;/a&gt; walks through the builder options and what each really costs.&lt;/p&gt;

&lt;h2&gt;
  
  
  Tech stack mistakes that cost startups a rebuild
&lt;/h2&gt;

&lt;p&gt;The expensive mistakes are almost never the wrong framework. They are architecture and ownership choices that look like speed in month one and become a rewrite in month nine.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Microservices from day one.&lt;/strong&gt; One deployable application is faster to build and debug. Split it when a real boundary forces you to.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Choosing for the CV, not the company.&lt;/strong&gt; A developer who picks a stack because they want to learn it is spending your runway on their education. Ask why, and listen for "hiring" and "handover" in the answer.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Building what you can buy.&lt;/strong&gt; Custom auth, a custom payment flow, a homemade admin panel before there is anything to administer. Each one is weeks of work no user sees.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Planning for scale you have not reached.&lt;/strong&gt; Sharding, multi-region, event streaming. If you get the traffic problem, it is the best problem you will ever have, and you will have the money to solve it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Vendors you cannot use in your market.&lt;/strong&gt; In regulated markets, check compliance at selection time. A healthtech MVP needs every vendor that touches patient data to sign a BAA; finding out during your first hospital review costs you the deal. In India, the &lt;a href="https://kunalvohra.com/blog/dpdp-act-explained" rel="noopener noreferrer"&gt;DPDP Act&lt;/a&gt; sets obligations on how personal data is protected, which narrows your vendor list.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Accounts in someone else's name.&lt;/strong&gt; The repository, the cloud account, the domain and the app store listings should belong to your company from day one, not to your agency or freelancer. It is the hardest mistake to unwind.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;There is one honest exception to "boring wins". On one of my ventures, an edtech platform, the existing stack capped runs in the thousands, so we built our own Linux OS from zero. It worked: the first full run handled 500,000 assessments in an hour. But that was the product's core problem, chosen deliberately, not a stack preference. Unusual technology is justified only when it is the thing you are selling. The &lt;a href="https://kunalvohra.com/blog/the-edtech-problem-nobody-would-touch" rel="noopener noreferrer"&gt;full case study&lt;/a&gt; explains why.&lt;/p&gt;

&lt;h2&gt;
  
  
  How to check a stack recommendation you can't evaluate
&lt;/h2&gt;

&lt;p&gt;Ask whoever proposes the stack five questions. Honest builders answer in plain language; vague answers are the warning sign.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;How many engineers in our hiring market know this stack?&lt;/strong&gt; You want "lots", with a reason.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;If you disappeared tomorrow, how long until a new engineer ships a change?&lt;/strong&gt; Days is the right answer. Weeks means the stack, the setup or the documentation is wrong.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What are we buying instead of building?&lt;/strong&gt; Auth, payments, email and hosting should all be on the list.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What will this cost per month at launch, and at 10x the users?&lt;/strong&gt; They should be able to give a rough range for both.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;What will we have to change first when we grow, and roughly when?&lt;/strong&gt; A builder who says "nothing" is either selling or has not thought about it.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Investors check the stack too; &lt;a href="https://kunalvohra.com/blog/ai-startup-technical-due-diligence" rel="noopener noreferrer"&gt;AI startup technical due diligence&lt;/a&gt; covers what they look for, and the &lt;a href="https://kunalvohra.com/blog/startup-security-checklist" rel="noopener noreferrer"&gt;startup security checklist&lt;/a&gt; covers the basics under any stack.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who should choose your MVP tech stack?
&lt;/h2&gt;

&lt;p&gt;Whoever will own the architecture should choose the stack, and that should be someone senior who is not paid by the hour to build whatever they pick. For most pre-seed founders that is a short engagement, not a full-time hire.&lt;/p&gt;

&lt;p&gt;A full-time CTO costs $180,000 to $260,000 a year in the US plus equity, which is a lot to pay for a decision that takes a week. If you have a builder lined up and want the stack and scope checked before you sign, a &lt;a href="https://kunalvohra.com/technical-advisor" rel="noopener noreferrer"&gt;technical advisor&lt;/a&gt; retainer ($750 a month for 15+ hours, held through December 2026) covers it. If you need someone to own the architecture, the build and the first hires, that is a &lt;a href="https://kunalvohra.com/fractional-cto" rel="noopener noreferrer"&gt;fractional CTO&lt;/a&gt; engagement ($2,500 a month for 50+ hours on the same terms); the &lt;a href="https://kunalvohra.com/fractional-cto-cost" rel="noopener noreferrer"&gt;fractional CTO cost guide&lt;/a&gt; shows market rates beside mine. And if you want the whole MVP built on a stack chosen for handover, my team does that at $30 a screen; run your scope through the &lt;a href="https://kunalvohra.com/mvp-cost-calculator" rel="noopener noreferrer"&gt;MVP cost calculator&lt;/a&gt; for a first range, and read &lt;a href="https://kunalvohra.com/blog/how-long-does-it-take-to-build-an-mvp" rel="noopener noreferrer"&gt;how long it takes to build an MVP&lt;/a&gt; for the timeline.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;What is the best tech stack for an MVP in 2026?&lt;/strong&gt; For most web products: TypeScript and React (often with Next.js) on the frontend, one backend in TypeScript or Python, managed Postgres, a managed auth provider, Stripe for payments and a managed hosting platform. Add one cross-platform mobile app (React Native or Flutter) only if users need it. The best stack is the one your future engineers already know.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Should I use no-code for my MVP?&lt;/strong&gt; If the goal is to test demand quickly and cheaply, often yes. Plan for platform lock-in, pricing that grows with your users, and a likely rebuild once you need something the platform does not support. If the product's value lives in custom logic or sensitive data, start with code.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What is the best tech stack for an AI MVP?&lt;/strong&gt; The same foundation plus one model provider behind a thin swappable seam, hosted APIs rather than your own GPUs, Postgres with pgvector if you need retrieval, cost logging on every model call, and a small eval set you run on every prompt change. Defer fine-tuning and multi-provider routing until usage data justifies them.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Do I need a native mobile app for my MVP?&lt;/strong&gt; Usually not. Start with a mobile-friendly web app if users can live with it, or build one cross-platform app for both stores.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Will I have to rebuild my MVP later?&lt;/strong&gt; Parts of it, and that is normal. A mainstream stack, one deployable application and clean ownership of the code mean you refactor in pieces instead of rewriting from scratch.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How do I choose a tech stack if I am not technical?&lt;/strong&gt; Ask the five questions above (hiring pool, handover time, what is bought, monthly cost now and at 10x, what changes first), or have a senior person review the proposal before you sign.&lt;/p&gt;

&lt;p&gt;If you are about to pick a stack, or someone has proposed one you cannot judge, &lt;a href="https://calendly.com/kunalvohra" rel="noopener noreferrer"&gt;book a 30-minute call&lt;/a&gt;. Bring the proposal, and I will tell you what I would keep, what I would buy instead of build, and what will cost you a rebuild. Or &lt;a href="https://kunalvohra.com/contact" rel="noopener noreferrer"&gt;get in touch&lt;/a&gt; if you would rather write first.&lt;/p&gt;

</description>
      <category>startup</category>
      <category>career</category>
      <category>leadership</category>
      <category>entrepreneurship</category>
    </item>
    <item>
      <title>PHP2Shell (CVE-2026-109618 and CVE-2026-109619): What Happened and How to Fix It, Step by Step</title>
      <dc:creator>Kunal Vohra</dc:creator>
      <pubDate>Fri, 23 Oct 2026 12:00:00 +0000</pubDate>
      <link>https://dev.to/kunalvohra/php2shell-cve-2026-109618-and-cve-2026-109619-what-happened-and-how-to-fix-it-step-by-step-5fn7</link>
      <guid>https://dev.to/kunalvohra/php2shell-cve-2026-109618-and-cve-2026-109619-what-happened-and-how-to-fix-it-step-by-step-5fn7</guid>
      <description>&lt;p&gt;PHP2Shell is the name for two critical vulnerabilities in the PHP runtime itself, CVE-2026-109618 and CVE-2026-109619, patched on 21 October 2026. Both can give an attacker remote code execution on a web server without logging in, and on Apache with mod_php a single request can trigger the attack before your application code even runs. The fix is to update to PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35 (or your vendor's patched package), then restart every PHP worker. If you can't do that today, there are interim settings that reduce the risk, but the PHP project is clear that updating is the only complete fix.&lt;/p&gt;

&lt;p&gt;I'm Kunal Vohra. I have an M.Tech in Cybersecurity from NIT Kurukshetra and I've been the technical co-founder of six startups, so I've been on the receiving end of more than one "patch everything by tonight" morning. This is the guide I'd hand a founder or a small engineering team: what happened in plain language, then the exact steps, in order.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is PHP2Shell?
&lt;/h2&gt;

&lt;p&gt;PHP2Shell is two memory-safety bugs in how PHP sets up and tears down web requests. Neither lives in WordPress, Laravel or your code; they are in the PHP engine, so every framework running on an affected version inherits them.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;CVE-2026-109618: remote code execution through out-of-memory corruption.&lt;/strong&gt; If PHP hits its memory limit at the wrong moment while setting up a request, some internal request data is left only partly initialised. The cleanup code then treats that broken data as valid, which corrupts memory. On Apache with mod_php, researchers demonstrated code execution with default settings from an unauthenticated request. On PHP-FPM it was confirmed through certain application code paths.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;CVE-2026-109619: remote code execution through stale global state.&lt;/strong&gt; A fatal error at the wrong point can leave internal state pointing at memory that has already been freed. The next request handled by the same worker can then use that leftover state, a classic use-after-free. Code execution was demonstrated against a real application on Apache with mod_php.&lt;/p&gt;

&lt;p&gt;The PHP project rates both &lt;strong&gt;critical, CVSS v4.0 9.2&lt;/strong&gt;. (A leaked CVE listing showed 10.0; that was wrong.) The researchers at Calif reported the issues through coordinated disclosure.&lt;/p&gt;

&lt;h2&gt;
  
  
  How the PHP2Shell disclosure unfolded
&lt;/h2&gt;

&lt;p&gt;The timeline matters because attackers had a head start, so "we'll patch in the next maintenance window" is not good enough here.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Before 20 October:&lt;/strong&gt; Calif reports the bugs to the PHP security team. Linux distributions and long-term-support vendors receive patches in advance.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;17 October:&lt;/strong&gt; Amazon Threat Intelligence records a PHP2Shell exploitation attempt on its MadPot honeypot, four days before public disclosure. How the attacker got the details is not known.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;20 October:&lt;/strong&gt; Details leak on social media ahead of the planned 22 October disclosure. The PHP team announces an emergency release for the next day.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;21 October, 14:00 UTC:&lt;/strong&gt; PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35 ship with the advisories. watchTowr reports seeing exploitation in its honeypot network from the same day.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If there is one line to remember from this incident: &lt;strong&gt;a single unauthenticated request, before any of your code runs, was enough to take over an Apache mod_php server.&lt;/strong&gt; No plugin choice or code review protects you from that. Only the runtime version does.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who is affected by CVE-2026-109618 and CVE-2026-109619?
&lt;/h2&gt;

&lt;p&gt;Effectively every PHP installation that serves web requests and hasn't been patched since 21 October. Google's advisory puts the affected range at PHP 5.4 through 8.5.11.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Affected server setups:&lt;/strong&gt; PHP-FPM, Apache with mod_php (highest risk), php-cgi, FrankenPHP and LiteSpeed.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Fixed versions by branch:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;PHP 8.5: fixed in &lt;strong&gt;8.5.12&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;PHP 8.4: fixed in &lt;strong&gt;8.4.27&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;PHP 8.3: fixed in &lt;strong&gt;8.3.36&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;PHP 8.2: fixed in &lt;strong&gt;8.2.35&lt;/strong&gt; (note that 8.2 reaches end of life on 31 December 2026)&lt;/li&gt;
&lt;li&gt;PHP 7.4, 8.0 and 8.1: no official release. The PHP team published minimal source backports, and some distributions and LTS vendors ship patched packages. Otherwise, upgrade.&lt;/li&gt;
&lt;li&gt;PHP 7.3 and older, including PHP 5: no fix at all. Upgrade to a supported branch.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;That covers most WordPress sites, most Laravel and Symfony apps, many internal admin tools, and a surprising number of appliances. Palo Alto Networks, for example, said all PAN-OS and Panorama versions are affected. If your startup's main product is in Node, Python or Go, you are probably still running PHP somewhere: the marketing site, a help centre, an old billing tool, a phpMyAdmin someone set up in year one. My own site runs on Next.js, so the site itself was never exposed, but that's exactly the kind of assumption the first step below exists to test.&lt;/p&gt;

&lt;h2&gt;
  
  
  How to fix PHP2Shell, step by step
&lt;/h2&gt;

&lt;p&gt;Work through these in order. Steps 1 to 4 are the fix; step 5 is for anything you can't patch today; steps 6 and 7 are what most teams skip.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 1: Find every PHP installation you run
&lt;/h3&gt;

&lt;p&gt;Make a list before you patch anything, starting with whatever is reachable from the internet. Include servers, containers, base images, staging environments, and anything a vendor hosts for you.&lt;/p&gt;

&lt;p&gt;On each Linux server, these show what is installed and what is running:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# Which PHP the command line uses&lt;/span&gt;
php &lt;span class="nt"&gt;-v&lt;/span&gt;

&lt;span class="c"&gt;# Installed PHP packages (Debian/Ubuntu)&lt;/span&gt;
dpkg &lt;span class="nt"&gt;-l&lt;/span&gt; | &lt;span class="nb"&gt;grep&lt;/span&gt; &lt;span class="nt"&gt;-i&lt;/span&gt; php

&lt;span class="c"&gt;# Installed PHP packages (RHEL, Rocky, Alma, Fedora)&lt;/span&gt;
rpm &lt;span class="nt"&gt;-qa&lt;/span&gt; | &lt;span class="nb"&gt;grep&lt;/span&gt; &lt;span class="nt"&gt;-i&lt;/span&gt; php

&lt;span class="c"&gt;# PHP processes actually running right now&lt;/span&gt;
ps &lt;span class="nt"&gt;-eo&lt;/span&gt; pid,lstart,cmd | &lt;span class="nb"&gt;grep&lt;/span&gt; &lt;span class="nt"&gt;-E&lt;/span&gt; &lt;span class="s2"&gt;"php|httpd|apache2"&lt;/span&gt; | &lt;span class="nb"&gt;grep&lt;/span&gt; &lt;span class="nt"&gt;-v&lt;/span&gt; &lt;span class="nb"&gt;grep&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Be careful with &lt;code&gt;php -v&lt;/code&gt;: it reports the command-line PHP, which can be a different build from the one serving web traffic. A server can have several PHP versions side by side. Check the FPM binary (for example &lt;code&gt;php-fpm8.3 -v&lt;/code&gt;) or the Apache module, and check every container image separately:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;docker run &lt;span class="nt"&gt;--rm&lt;/span&gt; &lt;span class="nt"&gt;--entrypoint&lt;/span&gt; php your-image:tag &lt;span class="nt"&gt;-v&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Step 2: Check whether each one is already patched
&lt;/h3&gt;

&lt;p&gt;Compare each version to the fixed list above. The catch: distributions often backport fixes &lt;strong&gt;without changing the upstream version number&lt;/strong&gt; , so an Ubuntu or Red Hat package can say 8.1 or 8.3.x and still be patched. For distribution packages, compare the full package version against your vendor's security notice (Ubuntu's is USN-8948-1) rather than the PHP version string.&lt;/p&gt;

&lt;p&gt;On managed hosting, ask your host directly whether the 21 October release is applied to your server. On WordPress you can see the PHP version under Tools, then Site Health, then Info, then Server, but the same backport caveat applies.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 3: Patch PHP
&lt;/h3&gt;

&lt;p&gt;Pick the path that matches each install.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Ubuntu and Debian servers:&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;sudo &lt;/span&gt;apt update
&lt;span class="nb"&gt;sudo &lt;/span&gt;apt &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;--only-upgrade&lt;/span&gt; &lt;span class="s1"&gt;'php*'&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;RHEL, Rocky, Alma and Fedora:&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;sudo &lt;/span&gt;dnf upgrade &lt;span class="s1"&gt;'php*'&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If you install PHP from the volunteer-run &lt;code&gt;rpms.remirepo.net&lt;/code&gt; repository, the PHP team advised switching to your distribution's packages or building from source, because that repository had no update on release day.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Docker and Kubernetes:&lt;/strong&gt; a running container never updates itself. Rebuild every image on a patched base and redeploy:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;docker build &lt;span class="nt"&gt;--pull&lt;/span&gt; &lt;span class="nt"&gt;--no-cache&lt;/span&gt; &lt;span class="nt"&gt;-t&lt;/span&gt; your-image:patched &lt;span class="nb"&gt;.&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then roll the new image out to every service and pod. On Google Cloud, App Engine standard on PHP 8.2+ and Cloud Run services with automatic base-image updates are patched for you; everything else needs a rebuild and redeploy.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Building from source:&lt;/strong&gt; download the fixed release from php.net, or apply the minimal backport branch for 7.4, 8.0 or 8.1.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Managed WordPress and shared hosting:&lt;/strong&gt; most hosts patched at server level or put a firewall rule in front while they rolled out the fix. Your job is to confirm it in writing, take a backup, and, if the host forces an upgrade from PHP 8.1 or older, test the site (especially checkout and forms) on a staging copy. Don't edit server settings yourself on shared hosting.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;PHP 8.1 or older with no vendor patch:&lt;/strong&gt; this is the moment to upgrade to 8.2 or newer. Test themes, plugins and libraries on staging first; old code is exactly where an upgrade breaks.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 4: Restart every PHP worker and confirm the new build is running
&lt;/h3&gt;

&lt;p&gt;This is the step people miss. Updating the package does nothing for workers that are already running; they keep the vulnerable code in memory until they restart.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# PHP-FPM (use your version's service name)&lt;/span&gt;
&lt;span class="nb"&gt;sudo &lt;/span&gt;systemctl restart php8.3-fpm

&lt;span class="c"&gt;# Apache with mod_php&lt;/span&gt;
&lt;span class="nb"&gt;sudo &lt;/span&gt;systemctl restart apache2 &lt;span class="c"&gt;# or: httpd&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then run the &lt;code&gt;ps -eo pid,lstart,cmd&lt;/code&gt; command from step 1 again and make sure no PHP or Apache process started before the patch. Finally, confirm the web-facing runtime, not just the CLI, reports the fixed version or patched package. If you use a temporary &lt;code&gt;phpinfo()&lt;/code&gt; page to check, delete it immediately afterwards.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 5: If you can't patch today, reduce the risk
&lt;/h3&gt;

&lt;p&gt;The PHP project published interim settings. They lower the risk; they do not make an unpatched server safe. Apply them in &lt;code&gt;php.ini&lt;/code&gt;, the FPM pool file or the Apache config (not with &lt;code&gt;ini_set()&lt;/code&gt; or &lt;code&gt;.user.ini&lt;/code&gt;, which apply too late), then restart every worker.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;For CVE-2026-109619 (stale global state):&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight ini"&gt;&lt;code&gt;&lt;span class="c"&gt;; Disable session upload progress
&lt;/span&gt;&lt;span class="py"&gt;session.upload_progress.enabled&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;=&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s"&gt;0&lt;/span&gt;
&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="c"&gt;; PHP 8.5 only: disable fatal error backtraces
&lt;/span&gt;&lt;span class="py"&gt;fatal_error_backtraces&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;=&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s"&gt;0&lt;/span&gt;
&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;For CVE-2026-109618 (out-of-memory corruption):&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight conf"&gt;&lt;code&gt;; &lt;span class="n"&gt;In&lt;/span&gt; &lt;span class="n"&gt;the&lt;/span&gt; &lt;span class="n"&gt;PHP&lt;/span&gt;-&lt;span class="n"&gt;FPM&lt;/span&gt; &lt;span class="n"&gt;pool&lt;/span&gt;, &lt;span class="n"&gt;so&lt;/span&gt; &lt;span class="n"&gt;applications&lt;/span&gt; &lt;span class="n"&gt;cannot&lt;/span&gt; &lt;span class="n"&gt;override&lt;/span&gt; &lt;span class="n"&gt;it&lt;/span&gt;
&lt;span class="n"&gt;php_admin_value&lt;/span&gt;[&lt;span class="n"&gt;memory_limit&lt;/span&gt;] = -&lt;span class="m"&gt;1&lt;/span&gt;

; &lt;span class="n"&gt;Keep&lt;/span&gt; &lt;span class="n"&gt;request&lt;/span&gt; &lt;span class="n"&gt;bodies&lt;/span&gt; &lt;span class="n"&gt;small&lt;/span&gt;. &lt;span class="n"&gt;If&lt;/span&gt; &lt;span class="n"&gt;you&lt;/span&gt; &lt;span class="n"&gt;don&lt;/span&gt;&lt;span class="err"&gt;'&lt;/span&gt;&lt;span class="n"&gt;t&lt;/span&gt; &lt;span class="n"&gt;need&lt;/span&gt; &lt;span class="n"&gt;uploads&lt;/span&gt;:
&lt;span class="n"&gt;post_max_size&lt;/span&gt; = &lt;span class="m"&gt;8&lt;/span&gt;&lt;span class="n"&gt;M&lt;/span&gt;
; &lt;span class="n"&gt;If&lt;/span&gt; &lt;span class="n"&gt;you&lt;/span&gt; &lt;span class="n"&gt;do&lt;/span&gt; &lt;span class="n"&gt;need&lt;/span&gt; &lt;span class="n"&gt;uploads&lt;/span&gt;:
; &lt;span class="n"&gt;post_max_size&lt;/span&gt; = &lt;span class="m"&gt;32&lt;/span&gt;&lt;span class="n"&gt;M&lt;/span&gt;
; &lt;span class="n"&gt;upload_max_filesize&lt;/span&gt; = &lt;span class="m"&gt;24&lt;/span&gt;&lt;span class="n"&gt;M&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Removing the memory limit is a deliberate trade: it takes away the out-of-memory trigger but lets one bad request use more memory, so put a container or cgroup memory limit around PHP as the safety net. The PHP team specifically warns against using &lt;code&gt;RLIMIT_AS&lt;/code&gt; or &lt;code&gt;RLIMIT_DATA&lt;/code&gt; for this, because they can trigger the vulnerable failure.&lt;/p&gt;

&lt;p&gt;Also:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Enforce request size, field count, header count and multipart limits at your web server or proxy, including for requests without a &lt;code&gt;Content-Length&lt;/code&gt; header. Block multipart requests entirely if the app doesn't take uploads.&lt;/li&gt;
&lt;li&gt;Remove invalid or deprecated per-request PHP settings (in &lt;code&gt;.htaccess&lt;/code&gt;, Apache &lt;code&gt;php_value&lt;/code&gt; lines, or &lt;code&gt;PHP_VALUE&lt;/code&gt; passed from a front end). Hiding the warnings doesn't help.&lt;/li&gt;
&lt;li&gt;On PHP-FPM, &lt;code&gt;pm.max_requests = 1&lt;/code&gt; recycles a worker after every request, which limits stale state, at a real performance cost.&lt;/li&gt;
&lt;li&gt;If a system can't be patched or mitigated, take it off the public internet until it can. The UK's NCSC advised exactly that.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  Step 6: Add a firewall rule in front, as a second layer
&lt;/h3&gt;

&lt;p&gt;A web application firewall buys time while you patch; it doesn't replace patching. Cloudflare deployed rules for the known exploit patterns to all customers, including free plans; on paid plans, check its managed rules are set to block, not just log. On Google Cloud, Cloud Armor's managed rules cover it; Google recommends enabling at least the File Upload and RFI categories and blocking request bodies too large to inspect. If you're on AWS, Azure or another WAF, check your provider's PHP2Shell advisory.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 7: Check whether you were already hit
&lt;/h3&gt;

&lt;p&gt;Exploitation started before the patch, so a server that was exposed between 17 October and the moment you patched should be checked, not assumed clean. None of the advisories I read publish specific indicators of compromise yet, so this is a general post-RCE check rather than a PHP2Shell signature:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;New or changed files in the web root,&lt;/strong&gt; especially &lt;code&gt;.php&lt;/code&gt; files in upload folders, cache folders or anywhere PHP shouldn't be:
&lt;/li&gt;
&lt;/ul&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;sudo &lt;/span&gt;find /var/www &lt;span class="nt"&gt;-name&lt;/span&gt; &lt;span class="s2"&gt;"*.php"&lt;/span&gt; &lt;span class="nt"&gt;-newermt&lt;/span&gt; &lt;span class="s2"&gt;"2026-10-17"&lt;/span&gt; &lt;span class="nt"&gt;-ls&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Unexpected processes or outbound connections&lt;/strong&gt; from the web server user (&lt;code&gt;www-data&lt;/code&gt;, &lt;code&gt;apache&lt;/code&gt;, &lt;code&gt;nginx&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;New cron jobs, SSH keys or system users,&lt;/strong&gt; and on WordPress, new administrator accounts or plugins you didn't install.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Web server logs&lt;/strong&gt; with unusually large or malformed requests, especially multipart uploads to endpoints that don't take uploads, around worker crashes.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Rotate secrets&lt;/strong&gt; the server could read (database passwords, API keys, cloud credentials in environment files) if anything above looks wrong. Code execution means anything the PHP process could read should be treated as exposed.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you find something, isolate the server, keep the logs, and rebuild from a known-good image rather than cleaning in place. Scanners can help with discovery: Searchlight Cyber released an open-source &lt;code&gt;php2shell-scanner&lt;/code&gt;, and the major vulnerability scanners added checks.&lt;/p&gt;

&lt;h3&gt;
  
  
  After the patch: undo the temporary settings
&lt;/h3&gt;

&lt;p&gt;Once every runtime is patched and restarted, restore a finite &lt;code&gt;memory_limit&lt;/code&gt;, revert &lt;code&gt;pm.max_requests&lt;/code&gt; and any upload limits you tightened beyond what the app needs, and test under load. Then rescan to confirm nothing was missed.&lt;/p&gt;

&lt;h2&gt;
  
  
  What PHP2Shell should change in how a startup runs security
&lt;/h2&gt;

&lt;p&gt;The patch is the easy part. The hard part, every time, is step 1: nobody had a list. Three habits make the next one of these a two-hour job instead of a two-day one.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Keep an inventory of runtimes, not just services.&lt;/strong&gt; For every system, write down the language runtime, version, where it's hosted and who owns it. The forgotten marketing site and the internal tool from year one are always the ones that get hit.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Never run an end-of-life runtime on the internet.&lt;/strong&gt; PHP 8.1 and older got no official fix. Anything past end of life means "the next critical bug has no patch," so build the upgrade into the roadmap before the deadline, not during an incident. PHP 8.2 joins that list on 31 December.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Make patch-and-restart boring.&lt;/strong&gt; Automatic base-image rebuilds, a deploy pipeline that can roll every service in an hour, and a written order of operations. This sits in the first section of my &lt;a href="https://kunalvohra.com/blog/startup-security-checklist" rel="noopener noreferrer"&gt;startup security checklist&lt;/a&gt;, and it's one of the first things investors probe in &lt;a href="https://kunalvohra.com/blog/ai-startup-technical-due-diligence" rel="noopener noreferrer"&gt;technical due diligence&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;If you don't have someone who owns this, it doesn't need to be a full-time CTO at $180,000 to $260,000 a year. A &lt;a href="https://kunalvohra.com/technical-advisor" rel="noopener noreferrer"&gt;technical advisor&lt;/a&gt; retainer ($750 a month for 15+ hours, held through December 2026) covers a review of your inventory, patch process and exposure. If you need someone to own security, architecture and the team, that's a &lt;a href="https://kunalvohra.com/fractional-cto" rel="noopener noreferrer"&gt;fractional CTO&lt;/a&gt; engagement ($2,500 a month for 50+ hours on the same terms); the &lt;a href="https://kunalvohra.com/fractional-cto-cost" rel="noopener noreferrer"&gt;fractional CTO cost guide&lt;/a&gt; shows market rates beside mine.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;What is PHP2Shell?&lt;/strong&gt; PHP2Shell is the name for two critical PHP runtime vulnerabilities, CVE-2026-109618 and CVE-2026-109619, disclosed and patched on 21 October 2026. Both can allow unauthenticated remote code execution on servers running PHP, rated CVSS v4.0 9.2.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Which PHP versions fix CVE-2026-109618 and CVE-2026-109619?&lt;/strong&gt; PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35. PHP 7.4, 8.0 and 8.1 have minimal source backports and some vendor packages; PHP 7.3 and older have no fix. Distribution packages may be patched under an older version number, so check your vendor's advisory.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is my WordPress site vulnerable to PHP2Shell?&lt;/strong&gt; If it runs on an unpatched PHP version, yes; the bugs are in PHP itself, not in WordPress or plugins. Most managed and shared hosts have patched or put firewall protection in place, but confirm with your host in writing that the 21 October release is applied to your server.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is updating PHP enough, or do I need to restart?&lt;/strong&gt; You need to restart. Running PHP-FPM workers and Apache processes keep the old code in memory until they restart, so restart every PHP service after updating and confirm no process predates the patch.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Can I mitigate PHP2Shell without updating?&lt;/strong&gt; Partly. Disabling session upload progress, removing the memory limit with a cgroup limit as a backstop, tightening request size limits, and putting a WAF in front all reduce risk. The PHP project says none of these make an unpatched installation safe; take unpatchable systems off the internet.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is PHP2Shell being exploited?&lt;/strong&gt; Yes. Amazon recorded an exploitation attempt on 17 October, before disclosure, and watchTowr reported exploitation in its honeypots from 21 October. Treat any server that was exposed in that window as worth checking for compromise.&lt;/p&gt;

&lt;p&gt;If you're not sure where PHP is running in your company, or whether what you found is clean, &lt;a href="https://calendly.com/kunalvohra" rel="noopener noreferrer"&gt;book a 30-minute call&lt;/a&gt;. Bring your list of servers and hosts, and I'll help you work out what to patch first and what to check. Or &lt;a href="https://kunalvohra.com/contact" rel="noopener noreferrer"&gt;get in touch&lt;/a&gt; if you'd rather write first.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://www.php.net/archive/2026.php#2026-10-21-5" rel="noopener noreferrer"&gt;PHP security release announcement, 21 October 2026&lt;/a&gt; (php.net)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://thephp.foundation/blog/2026/10/21/todays-php-security-release-and-how-were-strengthening-php-security/" rel="noopener noreferrer"&gt;Today's PHP Security Release&lt;/a&gt; (The PHP Foundation)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://watchtowr.com/intelligence/php-cve-2026-109618-faq/" rel="noopener noreferrer"&gt;PHP2Shell FAQ&lt;/a&gt; (watchTowr)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://blog.qualys.com/vulnerabilities-threat-research/2026/10/20/php-announces-critical-security-release-for-two-vulnerabilities" rel="noopener noreferrer"&gt;PHP2Shell: find, mitigate and patch&lt;/a&gt; (Qualys)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://cloud.google.com/blog/products/identity-security/responding-to-core-php-vulnerabilities-cve-2026-109618-cve-2026-109619" rel="noopener noreferrer"&gt;Responding to core PHP vulnerabilities&lt;/a&gt; (Google Cloud)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.helpnetsecurity.com/2026/10/21/php2shell-php-vulnerabilities-patched-cve-2026-109618-cve-2026-109619/" rel="noopener noreferrer"&gt;PHP2Shell patched, update now&lt;/a&gt; (Help Net Security)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.codeable.io/blog/php-vulnerability-cve-2026-109618/" rel="noopener noreferrer"&gt;What WordPress site owners need to do&lt;/a&gt; (Codeable)&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>security</category>
      <category>cybersecurity</category>
      <category>webdev</category>
      <category>startup</category>
    </item>
    <item>
      <title>Next.js Security Update (October 2026): CVE-2026-96889, the use cache RCE, and How to Fix Your App Step by Step</title>
      <dc:creator>Kunal Vohra</dc:creator>
      <pubDate>Fri, 23 Oct 2026 12:00:00 +0000</pubDate>
      <link>https://dev.to/kunalvohra/nextjs-security-update-october-2026-cve-2026-96889-the-use-cache-rce-and-how-to-fix-your-app-4o2p</link>
      <guid>https://dev.to/kunalvohra/nextjs-security-update-october-2026-cve-2026-96889-the-use-cache-rce-and-how-to-fix-your-app-4o2p</guid>
      <description>&lt;p&gt;The October 2026 Next.js security update fixes three vulnerabilities: two critical unauthenticated remote code execution flaws (GHSA-673q-mjm6-r5xr in &lt;code&gt;use cache&lt;/code&gt; functions, and CVE-2026-96889 in image optimization of SVG files through sharp) and a high-severity denial of service in Server Actions. The fix is to upgrade &lt;code&gt;next&lt;/code&gt; to &lt;strong&gt;16.4.1&lt;/strong&gt; or &lt;strong&gt;15.5.28&lt;/strong&gt; , make sure &lt;strong&gt;sharp is 0.35.5&lt;/strong&gt; , then rebuild and redeploy. Whether your app was actually exploitable depends on three things: your Next.js version, whether you process untrusted SVGs, and whether you use Server Actions. This guide checks all three, in order.&lt;/p&gt;

&lt;p&gt;I'm Kunal Vohra, a security researcher and cybersecurity advisor with an M.Tech in Cybersecurity, and this site runs on Next.js. So I didn't just read the advisories; I ran this exact checklist on kunalvohra.com this week. It was on Next.js 15.1.11, inside two of the three affected ranges. I'll show what I found and what I changed as we go.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the October 2026 Next.js vulnerabilities are
&lt;/h2&gt;

&lt;p&gt;Vercel published the fixes on 14 October 2026 as an out-of-band release, outside the normal release cycle. Two of the three issues score &lt;strong&gt;9.5 on CVSS v4.0&lt;/strong&gt; , and neither needs the attacker to log in.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Remote code execution in &lt;code&gt;use cache&lt;/code&gt; functions (GHSA-673q-mjm6-r5xr), critical.&lt;/strong&gt; Affects Next.js 16.0.0 through 16.4.0 apps using the App Router with Cache Components, where a &lt;code&gt;use cache&lt;/code&gt; function returns more than one typed array. React calls the exact pattern rare, but the advisory says there is no workaround and recommends upgrading even if you have no such function. The matching React packages (&lt;code&gt;react-server-dom-webpack&lt;/code&gt;, &lt;code&gt;-parcel&lt;/code&gt;, &lt;code&gt;-turbopack&lt;/code&gt;) are fixed in 19.0.9, 19.1.10, 19.2.9 and 19.3.1.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Remote code execution in image optimization (CVE-2026-96889, GHSA-2464-fhr4-w9xw), critical.&lt;/strong&gt; A memory bug in librsvg, the SVG library used underneath sharp, can be triggered when an attacker-supplied SVG is processed on glibc-based Linux. It affects Next.js 10.0.0 through 15.5.27 and 16.0.0 through 16.4.0. Next.js does not optimize SVGs by default, so the image optimizer is exposed only if you turned on &lt;code&gt;dangerouslyAllowSVG&lt;/code&gt;. &lt;code&gt;next/og&lt;/code&gt; running on the Node.js runtime also renders SVG through sharp, so an &lt;code&gt;ImageResponse&lt;/code&gt; that draws untrusted input is a second route in.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Denial of service in Server Actions (GHSA-p3m5-59mp-jv5r), high.&lt;/strong&gt; Crafted requests can pin one server process at full CPU and block everything else it serves. Affects Next.js 13.0.0 through 15.5.27 and 16.0.0 through 16.4.0, but only App Router apps with at least one Server Action. There is no workaround.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Fixed in:&lt;/strong&gt; Next.js &lt;strong&gt;16.4.1&lt;/strong&gt; (which ships sharp 0.35.5) and &lt;strong&gt;15.5.28&lt;/strong&gt; (which still needs sharp 0.35.5 forced; see step 4). Singapore's Cyber Security Agency issued an alert on 20 October telling teams to patch immediately. As of mid-October no exploitation had been reported, but the details are public now, and that window closes fast.&lt;/p&gt;

&lt;p&gt;The line I'd put in front of any engineering team this week: &lt;strong&gt;"npm audit is clean" did not mean "patched" for this release.&lt;/strong&gt; Two of the three advisories launched without CVE IDs and didn't show up in the npm advisory feed at first, and the sharp fix can be silently held back by your lockfile.&lt;/p&gt;

&lt;h2&gt;
  
  
  Is your Next.js app affected? A five-minute check
&lt;/h2&gt;

&lt;p&gt;Answer three questions. If every answer is "no", you were not exploitable, though you should still upgrade, because the next bug won't ask permission.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What version is actually installed?&lt;/strong&gt; Run this in the project, not your memory of &lt;code&gt;package.json&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm &lt;span class="nb"&gt;ls &lt;/span&gt;next sharp

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Anything below 15.5.28 on the 15 line, or 16.0.0 to 16.4.0, is in range for at least one issue. On 14.x and older there is no patched release in that line at all.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Do you process SVGs from anyone but yourself?&lt;/strong&gt; Search your config and code:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;grep&lt;/span&gt; &lt;span class="nt"&gt;-rn&lt;/span&gt; &lt;span class="s2"&gt;"dangerouslyAllowSVG"&lt;/span&gt; next.config.&lt;span class="k"&gt;*&lt;/span&gt;
&lt;span class="nb"&gt;grep&lt;/span&gt; &lt;span class="nt"&gt;-rln&lt;/span&gt; &lt;span class="s2"&gt;"next/og&lt;/span&gt;&lt;span class="se"&gt;\|&lt;/span&gt;&lt;span class="s2"&gt;ImageResponse"&lt;/span&gt; app src 2&amp;gt;/dev/null

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If &lt;code&gt;dangerouslyAllowSVG&lt;/code&gt; is on, or an &lt;code&gt;ImageResponse&lt;/code&gt; on the Node.js runtime draws user-supplied images or text, you are in scope for CVE-2026-96889 on glibc Linux (most Ubuntu and Debian servers and the default &lt;code&gt;node&lt;/code&gt; Docker images; Alpine uses musl and is outside the stated scope).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Do you use Server Actions or &lt;code&gt;use cache&lt;/code&gt;?&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nb"&gt;grep&lt;/span&gt; &lt;span class="nt"&gt;-rln&lt;/span&gt; &lt;span class="s2"&gt;"use server"&lt;/span&gt; app src 2&amp;gt;/dev/null
&lt;span class="nb"&gt;grep&lt;/span&gt; &lt;span class="nt"&gt;-rln&lt;/span&gt; &lt;span class="s2"&gt;"use cache"&lt;/span&gt; app src 2&amp;gt;/dev/null

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Any Server Action on an affected version means the denial of service applies. Any &lt;code&gt;use cache&lt;/code&gt; on 16.0 to 16.4 means upgrade now, full stop.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What I found on this site:&lt;/strong&gt; Next.js 15.1.11 and sharp 0.33.5, so in range for both the image RCE and the Server Actions DoS. But &lt;code&gt;dangerouslyAllowSVG&lt;/code&gt; was off, the Open Graph images only draw my own post titles and JPEG, PNG or GIF covers, and there are no Server Actions. So neither was reachable. One thing did bother me: the image config allowed remote images from &lt;strong&gt;any&lt;/strong&gt; host, which meant anyone could push files from anywhere through my server's image optimizer. Default SVG blocking stopped this bug, but that's one setting between you and the next one.&lt;/p&gt;

&lt;h2&gt;
  
  
  How to fix the Next.js vulnerabilities, step by step
&lt;/h2&gt;

&lt;p&gt;Seven steps. Do them on a branch, test locally, then deploy.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 1: Pick your target version
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;On &lt;strong&gt;16.x&lt;/strong&gt; : upgrade to &lt;strong&gt;16.4.1&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;On &lt;strong&gt;15.x&lt;/strong&gt; : upgrade to &lt;strong&gt;15.5.28&lt;/strong&gt; , the patched release on the 15 line. If you are on an early 15 release (I was on 15.1), this is a bigger jump than it sounds, so plan a proper test pass.&lt;/li&gt;
&lt;li&gt;On &lt;strong&gt;14.x or older&lt;/strong&gt; : there is no patched release in your line. Pin sharp to 0.35.5 now (step 4) and plan the move to 15.5.28 or 16.4.1. If you can't move soon, make sure nothing processes untrusted SVGs and that you have no Server Actions exposed.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  Step 2: Upgrade &lt;code&gt;next&lt;/code&gt; and its companions
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm &lt;span class="nb"&gt;install &lt;/span&gt;next@15.5.28 eslint-config-next@15.5.28
&lt;span class="c"&gt;# or, on the 16 line:&lt;/span&gt;
npm &lt;span class="nb"&gt;install &lt;/span&gt;next@16.4.1 eslint-config-next@16.4.1

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Keep &lt;code&gt;eslint-config-next&lt;/code&gt; on the same version as &lt;code&gt;next&lt;/code&gt;. If you use React Server Components outside Next.js, upgrade the &lt;code&gt;react-server-dom-*&lt;/code&gt; package instead.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 3: Read the upgrade notes for the versions you skipped
&lt;/h3&gt;

&lt;p&gt;On a minor-version jump, check the Next.js release notes between your version and the target for deprecations in config, caching defaults and middleware. Then run your type check:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx tsc &lt;span class="nt"&gt;--noEmit&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  Step 4: Force sharp to 0.35.5 (the step most 15.x apps miss)
&lt;/h3&gt;

&lt;p&gt;sharp &lt;strong&gt;0.35.5&lt;/strong&gt; is the only sharp release with the librsvg fix. Next.js 15 accepts older sharp versions, so an existing lockfile can keep a vulnerable sharp even after you upgrade &lt;code&gt;next&lt;/code&gt;. Other packages that depend on sharp (favicon generators, image tools) can pull in their own copy too. Force one version for the whole tree with an override in &lt;code&gt;package.json&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"dependencies"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"sharp"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"0.35.5"&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"overrides"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"sharp"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"0.35.5"&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;

&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;With pnpm, use &lt;code&gt;pnpm.overrides&lt;/code&gt;; with Yarn, &lt;code&gt;resolutions&lt;/code&gt;. Then reinstall and check that only one sharp is left:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm &lt;span class="nb"&gt;install
&lt;/span&gt;npm &lt;span class="nb"&gt;ls &lt;/span&gt;sharp

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Every line should say 0.35.5. sharp 0.35.5 needs Node.js 20.9 or newer, so check the Node version on your server, not just your laptop.&lt;/p&gt;

&lt;p&gt;If you can't upgrade today, the interim mitigation from the write-ups is to update sharp alone, or to block the SVG loader in your own sharp usage:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="nx"&gt;sharp&lt;/span&gt; &lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;sharp&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="nx"&gt;sharp&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;block&lt;/span&gt;&lt;span class="p"&gt;({&lt;/span&gt; &lt;span class="na"&gt;operation&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;VipsForeignLoadSvg&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="p"&gt;});&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That only covers code paths that use your sharp instance. Upgrading Next.js is still the fix.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 5: Tighten what your image optimizer will fetch
&lt;/h3&gt;

&lt;p&gt;This isn't in the advisory, but it's the change I'd make to every Next.js app this week. Replace wildcard remote patterns with the hosts you actually use:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="c1"&gt;// next.config.ts&lt;/span&gt;
&lt;span class="nx"&gt;images&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="nl"&gt;remotePatterns&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;
    &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="na"&gt;protocol&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;https&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="na"&gt;hostname&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;images.unsplash.com&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt; &lt;span class="p"&gt;},&lt;/span&gt;
  &lt;span class="p"&gt;],&lt;/span&gt;
&lt;span class="p"&gt;},&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Leave &lt;code&gt;dangerouslyAllowSVG&lt;/code&gt; off unless you truly need it, and if you do, only for SVGs you host yourself. On this site, after narrowing the list, a request for an image from any other host gets a 400 instead of being fetched and processed.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 6: Build, test, and check the running version
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm run build
npm run start

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Click through the pages that use images, Open Graph cards, forms and anything with Server Actions. Then confirm the server is actually running the new build. If you deploy with Docker, rebuild the image from the updated lockfile (&lt;code&gt;npm ci&lt;/code&gt; installs exactly what the lockfile says), and if you run PM2 or systemd, restart the process after deploying. On Vercel and Netlify, redeploy; Netlify also suggests deleting old deploy previews and branch deploys, which can stay vulnerable until they expire.&lt;/p&gt;

&lt;h3&gt;
  
  
  Step 7: Check whether anyone got there first
&lt;/h3&gt;

&lt;p&gt;No exploitation was reported in the first week, so this is a precaution rather than an emergency, unless you had &lt;code&gt;dangerouslyAllowSVG&lt;/code&gt; on or user-supplied images in &lt;code&gt;next/og&lt;/code&gt;. In that case:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Look in your access logs for requests to &lt;code&gt;/_next/image&lt;/code&gt; with SVG URLs, especially from hosts you don't recognise, and for bursts of worker crashes or restarts.&lt;/li&gt;
&lt;li&gt;For Server Actions, look for CPU spikes with a single request type behind them.&lt;/li&gt;
&lt;li&gt;If anything looks wrong, rotate the secrets the Node.js process could read (environment variables, API keys, database URLs) and redeploy from a clean build.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What I changed on kunalvohra.com
&lt;/h2&gt;

&lt;p&gt;For transparency, here is exactly what the fix looked like on a real production Next.js site this week:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;code&gt;next&lt;/code&gt; 15.1.11 to &lt;strong&gt;15.5.28&lt;/strong&gt; , &lt;code&gt;eslint-config-next&lt;/code&gt; to match.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;sharp&lt;/code&gt; pinned to &lt;strong&gt;0.35.5&lt;/strong&gt; as a direct dependency and an override. Before the override, a favicon generator in the dev dependencies would have kept its own older sharp.&lt;/li&gt;
&lt;li&gt;Image remote patterns narrowed from any host to &lt;code&gt;images.unsplash.com&lt;/code&gt;, the only remote host the site serves through the optimizer.&lt;/li&gt;
&lt;li&gt;Type check, full production build and a local smoke test (home page, blog posts, About, and a blocked image request from an unlisted host returning 400) before deploying.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Total time, including reading the advisories: under two hours. The longest part was confirming which conditions applied, which is why the five-minute check above comes first.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this release should change in how you run a Next.js app
&lt;/h2&gt;

&lt;p&gt;The patch is routine. The lessons are about process.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Don't treat &lt;code&gt;npm audit&lt;/code&gt; as your only signal.&lt;/strong&gt; Two of these advisories launched without CVEs and missed the npm feed at first. Subscribe to the Next.js and React security advisories on GitHub directly.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Know your transitive dependencies.&lt;/strong&gt; The image RCE wasn't in Next.js code at all; it was in librsvg, two layers down, under sharp. Your lockfile is part of your attack surface.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Keep the framework within a minor version or two of current.&lt;/strong&gt; Teams on 15.5 had a patch-version bump. Teams on 15.1 had a minor-version jump with release notes to read. Teams on 14 had no patch at all. Being far behind turns a one-hour fix into a migration project during an incident.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Allow-list by default.&lt;/strong&gt; Wildcard image hosts, open CORS and broad Server Action surfaces are fine until the day a bug lands in the code behind them. This sits in my &lt;a href="https://kunalvohra.com/blog/startup-security-checklist" rel="noopener noreferrer"&gt;startup security checklist&lt;/a&gt;, and it's the kind of thing a technically literate investor asks about in &lt;a href="https://kunalvohra.com/blog/ai-startup-technical-due-diligence" rel="noopener noreferrer"&gt;technical due diligence&lt;/a&gt;. The same week also brought a critical PHP runtime bug; my &lt;a href="https://kunalvohra.com/blog/php2shell-cve-2026-109618-how-to-fix" rel="noopener noreferrer"&gt;PHP2Shell fix guide&lt;/a&gt; covers that one.&lt;/p&gt;

&lt;p&gt;If nobody on your team owns this, it doesn't need a full-time CTO at $180,000 to $260,000 a year. A &lt;a href="https://kunalvohra.com/technical-advisor" rel="noopener noreferrer"&gt;technical advisor&lt;/a&gt; retainer ($750 a month for 15+ hours, held through December 2026) covers a dependency and exposure review like the one above. If you need someone to own security, architecture and the team, that's a &lt;a href="https://kunalvohra.com/fractional-cto" rel="noopener noreferrer"&gt;fractional CTO&lt;/a&gt; engagement ($2,500 a month for 50+ hours on the same terms); the &lt;a href="https://kunalvohra.com/fractional-cto-cost" rel="noopener noreferrer"&gt;fractional CTO cost guide&lt;/a&gt; shows market rates beside mine.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;What is CVE-2026-96889?&lt;/strong&gt; A critical (CVSS v4.0 9.5) unauthenticated remote code execution flaw in Next.js image optimization, caused by a memory bug in librsvg used through sharp when processing SVG files on glibc-based Linux. It affects Next.js 10.0.0 through 15.5.27 and 16.0.0 through 16.4.0, and is fixed in 15.5.28 and 16.4.1 with sharp 0.35.5.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Which Next.js versions fix the October 2026 vulnerabilities?&lt;/strong&gt; Next.js 16.4.1 and 15.5.28. On 15.5.28 you also need sharp 0.35.5, which you may have to force with an override. There is no patched release for 14.x or older.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is my Next.js app vulnerable if I don't use SVGs?&lt;/strong&gt; Not to CVE-2026-96889 through the image optimizer, which blocks SVGs unless &lt;code&gt;dangerouslyAllowSVG&lt;/code&gt; is on. Check &lt;code&gt;next/og&lt;/code&gt; image routes that render user-supplied content, and still check Server Actions (denial of service) and, on 16.x, &lt;code&gt;use cache&lt;/code&gt; (RCE). Upgrade either way.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Does &lt;code&gt;npm audit&lt;/code&gt; detect these Next.js vulnerabilities?&lt;/strong&gt; Not reliably at first. Two of the three advisories launched without CVE IDs and weren't in the npm advisory feed. Check versions directly with &lt;code&gt;npm ls next sharp&lt;/code&gt; and compare them to the fixed versions.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Are Vercel and Netlify apps affected?&lt;/strong&gt; The &lt;code&gt;use cache&lt;/code&gt; RCE affects any host if your app meets the conditions. Netlify says its image CDN doesn't use the vulnerable optimizer path and rates the other two as minimal impact on its platform; self-hosted apps on glibc Linux carry the most risk. Either way, upgrade and redeploy, including old previews.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is there a workaround if I can't upgrade?&lt;/strong&gt; For the image flaw, update sharp to 0.35.5 or block its SVG loader. For the &lt;code&gt;use cache&lt;/code&gt; RCE and the Server Actions denial of service, there is no workaround; upgrading is the only fix.&lt;/p&gt;

&lt;p&gt;If you're running Next.js and aren't sure what your app exposes, &lt;a href="https://calendly.com/kunalvohra" rel="noopener noreferrer"&gt;book a 30-minute call&lt;/a&gt;. Bring your &lt;code&gt;package.json&lt;/code&gt; and &lt;code&gt;next.config&lt;/code&gt;, and I'll tell you what applies, what to upgrade first, and what to lock down while you're there. Or &lt;a href="https://kunalvohra.com/contact" rel="noopener noreferrer"&gt;get in touch&lt;/a&gt; if you'd rather write first.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://github.com/vercel/next.js/security/advisories/GHSA-673q-mjm6-r5xr" rel="noopener noreferrer"&gt;GHSA-673q-mjm6-r5xr: unauthenticated RCE in &lt;code&gt;use cache&lt;/code&gt; functions&lt;/a&gt; (GitHub, vercel/next.js)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.csa.gov.sg/alerts-and-advisories/alerts/al-2026-148/" rel="noopener noreferrer"&gt;Critical Vulnerabilities in Next.js, AL-2026-148&lt;/a&gt; (Cyber Security Agency of Singapore)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.netlify.com/changelog/2026-10-14-nextjs-react-security-vulnerabilities/" rel="noopener noreferrer"&gt;Next.js and React security release, October 2026&lt;/a&gt; (Netlify)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://thehackernews.com/2026/10/nextjs-ships-out-of-band-patch-for-two.html" rel="noopener noreferrer"&gt;Next.js ships out-of-band patch for two critical RCE flaws&lt;/a&gt; (The Hacker News)&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://hol.org/blog/cve-2026-96889-nextjs-october-2026-security-update-rce" rel="noopener noreferrer"&gt;Next.js 16.4.1 and 15.5.28 fix two unauthenticated RCEs&lt;/a&gt; (HOL)&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>security</category>
      <category>cybersecurity</category>
      <category>webdev</category>
      <category>startup</category>
    </item>
    <item>
      <title>AI Agent Security for Startups in 2026: How to Ship Agents That Can't Be Turned Against You</title>
      <dc:creator>Kunal Vohra</dc:creator>
      <pubDate>Wed, 21 Oct 2026 12:00:00 +0000</pubDate>
      <link>https://dev.to/kunalvohra/ai-agent-security-for-startups-in-2026-how-to-ship-agents-that-cant-be-turned-against-you-5e7h</link>
      <guid>https://dev.to/kunalvohra/ai-agent-security-for-startups-in-2026-how-to-ship-agents-that-cant-be-turned-against-you-5e7h</guid>
      <description>&lt;p&gt;AI agent security comes down to one idea: an agent is only as dangerous as the tools, data and credentials you give it, and anything it reads can try to take control of it. A chatbot that answers questions can embarrass you. An agent that can send email, query your database, issue refunds or push code can be talked into doing those things for someone else. You can't make the model itself trustworthy, so you secure what sits around it: narrow permissions, server-side checks, a human approval step for anything that can't be undone, and logs you can actually read.&lt;/p&gt;

&lt;p&gt;I'm Kunal Vohra. I have an M.Tech in Cybersecurity from NIT Kurukshetra, I've published on prompt injection, and I run ColadAI, a multi-LLM platform, so I make these calls on my own products before I recommend them to anyone. This is the version of agent security that fits inside a startup's first launch.&lt;/p&gt;

&lt;h2&gt;
  
  
  What AI agent security actually means
&lt;/h2&gt;

&lt;p&gt;An AI agent is a model that can take actions: call tools, hit APIs, read and write data, and decide its own next step. AI agent security is the work of making sure those actions only happen for the right user, on the right data, within limits you set, even when the model is confused, wrong or deliberately manipulated.&lt;/p&gt;

&lt;p&gt;That framing matters because it moves the problem out of the prompt. The most common mistake I see is treating the system prompt as the security boundary: "You must never reveal customer data" or "Only refund orders under $50." Those lines are instructions, not controls. A model follows them most of the time, and "most of the time" is not a security property. The boundary has to live in code the model can't talk to.&lt;/p&gt;

&lt;p&gt;The sentence I'd put on the wall of every AI team: &lt;strong&gt;the blast radius of an AI agent is the sum of its permissions, not the quality of its prompt.&lt;/strong&gt; If the worst thing an agent can do is draft a reply for a human to approve, prompt injection is a nuisance. If it can move money, it's an incident.&lt;/p&gt;

&lt;h2&gt;
  
  
  The five AI agent security risks that matter for a startup
&lt;/h2&gt;

&lt;p&gt;For a startup shipping its first agent, five risks cover nearly everything I check, and each has a concrete fix.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Prompt injection, direct and indirect.&lt;/strong&gt; Direct injection is a user typing instructions to override yours. Indirect injection is worse for agents: the instructions arrive inside content the agent reads on someone else's behalf, such as an email, a web page, a support ticket, a PDF or a retrieved document. The agent can't reliably tell your instructions from the attacker's, because to the model it's all text. I covered the mechanics in &lt;a href="https://medium.kunalvohra.com/prompt-injection-attacks-explained-the-security-vulnerability-every-ai-builder-needs-to-understand-d3d17b8c1c0c" rel="noopener noreferrer"&gt;Prompt Injection Attacks Explained&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Excessive agency.&lt;/strong&gt; The agent has more tools, broader scopes or stronger credentials than its job needs. It is usually a week-one convenience ("just use the admin key for now") that turns every other risk here from a bug into a breach.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. Data exfiltration.&lt;/strong&gt; An injected instruction gets the agent to send data somewhere it shouldn't: into an outgoing email, a URL it fetches, a link it renders, or a tool call to an external service. It only needs one tool that can send data out and one piece of content that tells it to.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Confused deputy and broken tenancy.&lt;/strong&gt; The agent acts with its own service-level permissions instead of the permissions of the user it's working for. A user who can't see another customer's records asks the agent, and the agent, holding a key that can see everything, helpfully fetches them. In a multi-tenant B2B product this is the risk that ends enterprise deals.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. Runaway actions and runaway cost.&lt;/strong&gt; Agents loop. A bad plan, a tool that keeps returning errors, or a malicious input can trigger hundreds of calls: emails sent, records updated, tokens spent. Abuse of an AI endpoint often shows up first as a bill, and an agent can run up that bill on its own.&lt;/p&gt;

&lt;p&gt;Notice that only the first risk is specific to language models. The other four are old security problems (least privilege, egress control, authorisation, rate limiting) that an agent makes easier to trigger. That's good news: you already know how to fix them.&lt;/p&gt;

&lt;h2&gt;
  
  
  Securing AI agents: the permission model I start from
&lt;/h2&gt;

&lt;p&gt;The fastest way to make an agent safe is to decide, tool by tool, what it can do without asking, what it can do with approval, and what it can never do. I sort every tool into three buckets before any code is written.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Read-only tools the agent can call freely.&lt;/strong&gt; Searching your own knowledge base, reading the current user's own records, checking an order status. Even here, the query runs with the user's permissions, and results are filtered on the server.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Write tools that need a human confirmation.&lt;/strong&gt; Sending an email or message to a third party, updating a customer record, creating a ticket in someone else's queue. The agent prepares the action; a person sees exactly what will happen and clicks approve.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Actions the agent never takes.&lt;/strong&gt; Deleting data, moving money above a small fixed limit, changing permissions or credentials, deploying code to production, anything legal or contractual. If the business genuinely needs one of these automated later, it gets its own design review, not a new line in the tool list.&lt;/p&gt;

&lt;p&gt;Then four rules apply to every bucket:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Act as the user, not as the system.&lt;/strong&gt; Pass the end user's identity and scopes through to every tool call, and enforce authorisation in the tool, the way you would for any API request. The agent should never be able to see more than the person it's helping.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;One agent, one credential, minimum scope.&lt;/strong&gt; Give each agent its own API keys and service accounts with only the permissions its tools need. Never reuse a human admin's token. When something goes wrong, you want to revoke one key without taking the product down.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Validate every tool call in code.&lt;/strong&gt; The model proposes arguments; your code checks them. Is this order ID owned by this user? Is this refund under the limit? Is this email address on an allowed domain? Reject anything that fails, and log it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Make tools narrow.&lt;/strong&gt;"send_refund(order_id)" with a server-side limit is safer than "call_payments_api(endpoint, body)". Every general-purpose tool you expose is a general-purpose tool an attacker can use.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Prompt injection in AI agents: what actually helps
&lt;/h2&gt;

&lt;p&gt;You can't fully prevent prompt injection today, so design as if it will sometimes succeed and make sure success doesn't matter much. Filters and clever system prompts reduce the rate; they don't close the hole.&lt;/p&gt;

&lt;p&gt;What does help, in the order I'd apply it:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Separate untrusted content from instructions.&lt;/strong&gt; Mark retrieved documents, emails and web pages clearly as data in the prompt, and never let untrusted text sit in the same position as your instructions. This lowers the success rate; it is not a guarantee, so the steps below still apply.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Don't let one step both read untrusted content and take a sensitive action.&lt;/strong&gt; If an agent reads an inbound email, the step that reads it shouldn't also hold the tool that can forward your customer list. Split the work so the part exposed to untrusted input has the weakest tools.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Control egress.&lt;/strong&gt; Limit which domains the agent can fetch from or send to, strip or neutralise links and images in model output where the product doesn't need them, and treat any tool that sends data outside your system as a write action that needs approval.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Watch for the tell.&lt;/strong&gt; Injected instructions often show up as tool calls that don't match the user's request: a summarisation task that suddenly wants to send an email. Flag tool calls that fall outside the expected pattern for that task and route them to a human.&lt;/p&gt;

&lt;p&gt;Retrieval is an injection route too. If your agent uses RAG, a poisoned document in the index is as dangerous as a malicious email. I wrote about keeping retrieval tenant-safe in &lt;a href="https://kunalvohra.com/blog/rag-vs-fine-tuning" rel="noopener noreferrer"&gt;RAG vs fine-tuning for startups&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Human in the loop for AI agents: where approval earns its friction
&lt;/h2&gt;

&lt;p&gt;Put a human approval step on any action that is irreversible, external or financial, and nowhere else. Approval on everything trains users to click "yes" without reading, which is worse than no approval at all.&lt;/p&gt;

&lt;p&gt;A good approval step shows the exact action, not the agent's summary of it: the recipient, the full message, the amount, the record and the field being changed. It defaults to "no", it expires, and it's logged with who approved what, because "the AI did it" is not an answer a customer accepts.&lt;/p&gt;

&lt;p&gt;The trade-off founders worry about is that approvals make the product feel less magical. In my experience the opposite holds for B2B buyers: an agent that drafts and waits is easier to sell than one that acts alone, because the buyer's security team has to sign off on it. You can remove approvals for specific actions later, once logs show the agent gets them right, and that is a much easier conversation than adding them after an incident.&lt;/p&gt;

&lt;h2&gt;
  
  
  Logging, monitoring and limits: knowing what your agent did
&lt;/h2&gt;

&lt;p&gt;If you can't reconstruct what an agent did and why, you can't secure it, debug it or defend it in a security review. Log every tool call with the user, the arguments, the result, and the conversation turn that led to it, and keep personal data out of those logs unless the feature needs it.&lt;/p&gt;

&lt;p&gt;Then add the limits that stop a bad run from becoming a bad week:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Step and loop limits&lt;/strong&gt; on every agent run, so a confused plan stops after a fixed number of tool calls instead of running all night.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Rate limits per user and per tool,&lt;/strong&gt; especially on anything that sends messages or writes data.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hard spend limits on model keys,&lt;/strong&gt; set at the provider and alerted on in your own system. My &lt;a href="https://kunalvohra.com/blog/gemini-vs-claude-vs-grok-vs-chatgpt-api-for-business" rel="noopener noreferrer"&gt;comparison of the major AI APIs&lt;/a&gt; covers where each provider lets you set them.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A kill switch.&lt;/strong&gt; One flag that disables an agent's write tools without a deploy.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;These logs double as your evidence. Investors doing &lt;a href="https://kunalvohra.com/blog/ai-startup-technical-due-diligence" rel="noopener noreferrer"&gt;technical due diligence on an AI startup&lt;/a&gt; now ask what stops a malicious input from making the product take an action it shouldn't, and "here is the permission model, here are the approval rules, here are the logs" is the strong answer.&lt;/p&gt;

&lt;h2&gt;
  
  
  An AI agent security checklist for your first launch
&lt;/h2&gt;

&lt;p&gt;Before an agent goes in front of real users, I want a yes to each of these ten:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Every tool is listed, with its bucket (free, approval, never) written down.&lt;/li&gt;
&lt;li&gt;Tool calls run with the end user's identity and permissions, enforced in code.&lt;/li&gt;
&lt;li&gt;The agent has its own credentials, scoped to the minimum, and none of them belong to a human admin.&lt;/li&gt;
&lt;li&gt;Every tool argument is validated on the server before anything executes.&lt;/li&gt;
&lt;li&gt;Untrusted content (emails, web pages, documents, user uploads) is treated as data, and the step that reads it doesn't hold sensitive tools.&lt;/li&gt;
&lt;li&gt;Outbound data paths (email, HTTP fetches, links in output) are limited and reviewed.&lt;/li&gt;
&lt;li&gt;Irreversible, external and financial actions need a human approval that shows the exact action.&lt;/li&gt;
&lt;li&gt;Runs have step limits; users and tools have rate limits; model keys have spend limits.&lt;/li&gt;
&lt;li&gt;Every tool call is logged with user, arguments, result and approval, without unnecessary personal data.&lt;/li&gt;
&lt;li&gt;There's a kill switch for write tools, and someone knows how to use it.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;This sits on top of the basics, not instead of them. Identity, secrets, backups and the rest come first; my &lt;a href="https://kunalvohra.com/blog/startup-security-checklist" rel="noopener noreferrer"&gt;startup security checklist&lt;/a&gt; has them in the order I fix them, with a shorter AI section that this post expands. For a wider view of how attackers are using AI themselves, see &lt;a href="https://kunalvohra.com/blog/can-ai-hack-systems" rel="noopener noreferrer"&gt;Can AI Hack Systems?&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who should own AI agent security at a startup?
&lt;/h2&gt;

&lt;p&gt;Whoever owns the architecture should own agent security, because almost every fix is an architecture decision: which tools exist, what credentials they hold, where approvals sit.&lt;/p&gt;

&lt;p&gt;For most early AI startups that doesn't need to be a full-time CTO at $180,000 to $260,000 a year plus equity. If you have a team building the agent and want the design checked, a &lt;a href="https://kunalvohra.com/technical-advisor" rel="noopener noreferrer"&gt;technical advisor&lt;/a&gt; retainer ($750 a month for 15+ hours, held through December 2026) covers reviewing the tool list, the permission model and the approval rules before launch. If you need someone to own the AI roadmap, the security posture and the build, that is a &lt;a href="https://kunalvohra.com/fractional-cto" rel="noopener noreferrer"&gt;fractional CTO&lt;/a&gt; engagement ($2,500 a month for 50+ hours on the same terms); the &lt;a href="https://kunalvohra.com/fractional-cto-cost" rel="noopener noreferrer"&gt;fractional CTO cost guide&lt;/a&gt; shows market rates beside mine. And if you're still at the build stage, my team and I &lt;a href="https://kunalvohra.com/mvp-development/ai-startups" rel="noopener noreferrer"&gt;build MVPs for AI startups&lt;/a&gt; with scoped tools, approvals and cost logging in from the first week; the &lt;a href="https://kunalvohra.com/mvp-cost-calculator" rel="noopener noreferrer"&gt;MVP cost calculator&lt;/a&gt; gives a first range, including the security review option.&lt;/p&gt;

&lt;h2&gt;
  
  
  Frequently asked questions
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;What is AI agent security?&lt;/strong&gt; It is the practice of making sure an AI agent's actions (tool calls, API requests, data reads and writes) only happen for the right user, on the right data, within limits you set, even if the model is wrong or manipulated. The controls live in code around the model: permissions, validation, approvals, limits and logs.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What are the biggest AI agent security risks?&lt;/strong&gt; For a startup, five: prompt injection (especially indirect injection through content the agent reads), excessive agency from over-broad tools and credentials, data exfiltration through any tool that sends data out, confused-deputy access where the agent uses its own permissions instead of the user's, and runaway actions or cost from loops.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Can you prevent prompt injection in an AI agent?&lt;/strong&gt; Not completely, with today's models. You can lower the success rate by separating untrusted content from instructions, and you can make success harmless by keeping sensitive tools away from steps that read untrusted input, controlling outbound data paths and requiring approval for risky actions.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;When does an AI agent need a human in the loop?&lt;/strong&gt; For any action that is irreversible, reaches outside your system, or moves money: sending external messages, deleting or bulk-changing data, payments, permission changes and deployments. Read-only actions on the user's own data usually don't need it, and approvals on everything teach users to stop reading them.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is the system prompt a security control?&lt;/strong&gt; No. Instructions in the system prompt shape behaviour most of the time, but a determined input can override them. Treat them as guidance and enforce every real rule (who can see what, what limits apply) in server-side code the model can't change.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Do investors ask about AI agent security?&lt;/strong&gt; Increasingly, yes. A technically literate diligence partner will ask what stops a malicious input from making your product take an action it shouldn't. A written permission model, approval rules and tool-call logs make that a short conversation.&lt;/p&gt;

&lt;p&gt;If you're about to give an agent real tools, &lt;a href="https://calendly.com/kunalvohra" rel="noopener noreferrer"&gt;book a 30-minute call&lt;/a&gt;. Bring the list of tools you plan to expose, and I'll tell you which ones need approvals, which need narrowing, and which shouldn't exist yet. Or &lt;a href="https://kunalvohra.com/contact" rel="noopener noreferrer"&gt;get in touch&lt;/a&gt; if you'd rather write first.&lt;/p&gt;

</description>
      <category>security</category>
      <category>cybersecurity</category>
      <category>webdev</category>
      <category>startup</category>
    </item>
  </channel>
</rss>
