<?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: ash the coder</title>
    <description>The latest articles on DEV Community by ash the coder (@ashthecoder_000005).</description>
    <link>https://dev.to/ashthecoder_000005</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%2F4186206%2F30eddda5-c3b1-43e9-b05c-6ad6aa829a1e.png</url>
      <title>DEV Community: ash the coder</title>
      <link>https://dev.to/ashthecoder_000005</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/ashthecoder_000005"/>
    <language>en</language>
    <item>
      <title>Your Next Customer Is a Bot. Is Your Store Readable?</title>
      <dc:creator>ash the coder</dc:creator>
      <pubDate>Mon, 02 Nov 2026 10:05:21 +0000</pubDate>
      <link>https://dev.to/ashthecoder_000005/your-next-customer-is-a-bot-is-your-store-readable-5b45</link>
      <guid>https://dev.to/ashthecoder_000005/your-next-customer-is-a-bot-is-your-store-readable-5b45</guid>
      <description>&lt;p&gt;The Wall Street Journal just framed a fight I didn't have on my bingo card: shopping bots coming for Amazon. From the opening of their piece (Isabella Simonetti and Suzanne Vranica, Nov 1, 2026): Meta's new Muse shopping assistant is pitching itself as the place you start a purchase, which is awkward for Amazon's ad-supported discovery model. This post takes the headline and publicly visible opening as a starting point, not a full summary of their argument.&lt;/p&gt;

&lt;p&gt;Worth reading the original. What I want to add is the developer's version, because under the platform drama there's a practical question: what happens to a storefront when some of its "shoppers" are software acting for a human buyer?&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdrive.google.com%2Fuc%3Fexport%3Ddownload%26id%3D1vu5MHf1Y2neFazdlES5cnImipCel2xtP" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdrive.google.com%2Fuc%3Fexport%3Ddownload%26id%3D1vu5MHf1Y2neFazdlES5cnImipCel2xtP" alt="Side-by-side comparison: the human reader skims photos and reviews, notices badges and position, and can be swayed by presentation, which is the attention ads compete for; the agent, acting for the human, weighs specs, price, delivery and return policy, may drive a browser or use protocols like ACP, and can compare available product data. Caption: nobody knows how big agent shopping gets, but consistent, honest data works for both readers either way." width="1270" height="709"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Agents are starting to complete purchases, with caveats
&lt;/h2&gt;

&lt;p&gt;This is early and limited, but it's real:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;In September 2025, OpenAI turned on Instant Checkout inside ChatGPT for US users, starting with Etsy sellers, and, with Stripe, published the &lt;a href="https://openai.com/index/buy-it-in-chatgpt/" rel="noopener noreferrer"&gt;Agentic Commerce Protocol&lt;/a&gt; as an &lt;a href="https://stripe.com/newsroom/news/stripe-openai-instant-checkout" rel="noopener noreferrer"&gt;open standard&lt;/a&gt; - one defined way for a chatbot to go from "I want this" to a paid order.&lt;/li&gt;
&lt;li&gt;Amazon has been piloting "&lt;a href="https://www.aboutamazon.com/news/retail/amazon-shopping-app-buy-for-me-brands" rel="noopener noreferrer"&gt;Buy for Me&lt;/a&gt;" since April 2025: for some products, its agent goes to a brand's own website and fills in the checkout form for you.&lt;/li&gt;
&lt;li&gt;Google announced &lt;a href="https://blog.google/products-and-platforms/products/shopping/google-shopping-ai-mode-virtual-try-on-update/" rel="noopener noreferrer"&gt;agentic checkout in AI Mode&lt;/a&gt; at I/O in May 2025 and has been rolling conversational shopping out since.&lt;/li&gt;
&lt;li&gt;Perplexity shipped "&lt;a href="https://www.perplexity.ai/hub/blog/shop-like-a-pro" rel="noopener noreferrer"&gt;Buy with Pro&lt;/a&gt;" back in November 2024.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Note the pattern in the fine print: pilots, select merchants, specific countries. Agents aren't doing everyone's shopping. But the direction of travel is that the buyer's assistant increasingly can transact, not just recommend.&lt;/p&gt;

&lt;h2&gt;
  
  
  The ad model question, honestly
&lt;/h2&gt;

&lt;p&gt;Amazon's sponsored listings work because a human scans a results page and their attention is steerable - position one, a badge, a photo. If a growing share of purchase research runs through an agent, that slot machine has a problem: an agent working for the buyer doesn't have to rank results the way the platform's ad business prefers.&lt;/p&gt;

&lt;p&gt;But let's not write the obituary yet. Agents still act for humans, plenty of them work by driving an actual browser, and nothing stops the agent providers themselves from monetizing recommendations - sponsored answers are an obvious next experiment for every one of these companies. My honest read: the ad model gets renegotiated, not abolished. Whoever owns the agent owns a very valuable shelf, and shelves tend to end up for sale. That's a forecast, take it as opinion.&lt;/p&gt;

&lt;h2&gt;
  
  
  What an agent-friendly store actually looks like
&lt;/h2&gt;

&lt;p&gt;Here's where I think developers and small merchants have real, controllable work - no matter how the platform fight ends:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Data consistency.&lt;/strong&gt; Same price, stock and specs in your feed, your page and your checkout. Mismatches risk confusion or a failed checkout, whether the reader is a human, a bot or a price-comparison plugin.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Structured, honest product data.&lt;/strong&gt; Schema markup and clean attributes help machines understand your catalog. It's not a storefront by itself and guarantees nothing, but it's the part you fully control.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Readable policies.&lt;/strong&gt; Shipping, returns, warranty in plain, findable text. It avoids confusion for whatever reads your site - an agent, a search crawler, a careful human.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Standards where they fit.&lt;/strong&gt; If you sell through a channel that supports something like ACP, it's one integration option worth watching. It's not a universal requirement, and most stores don't need to think about it yet.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Notice what's not on the list: tricking the agents. Don't. Your checkout security, captchas and bot defenses are doing their job - the goal is to be legible to legitimate assistants your customers chose, not to let anything with a user agent walk in.&lt;/p&gt;

&lt;h2&gt;
  
  
  The builder's take
&lt;/h2&gt;

&lt;p&gt;Nobody knows how big agent shopping gets, and anyone who promises you the winner is selling something. What I'd bet on is narrower: purchase decisions made further from your product page mean your product information travels further without you. The stores that hold up are the ones whose data says the same thing everywhere, because they don't depend on a distracted human skimming the pretty version.&lt;/p&gt;

&lt;p&gt;That's not a revolution checklist. It's data hygiene with a deadline. Probably.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Prompted by &lt;a href="https://www.wsj.com/tech/the-shopping-bots-are-coming-for-amazon-701caa57" rel="noopener noreferrer"&gt;The Shopping Bots Are Coming for Amazon&lt;/a&gt; in the Wall Street Journal - this post takes that publicly visible opening as a starting point, and the framing above is my own take, not their reporting. Product announcements linked above are from the companies themselves, with their own geographic and merchant limitations.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>programming</category>
      <category>ecommerce</category>
      <category>agents</category>
      <category>ai</category>
    </item>
    <item>
      <title>Is Vibe Manufacturing the Next Vibe Coding?</title>
      <dc:creator>ash the coder</dc:creator>
      <pubDate>Mon, 02 Nov 2026 09:26:25 +0000</pubDate>
      <link>https://dev.to/ashthecoder_000005/vibe-manufacturing-is-vibe-coding-with-a-pricier-test-suite-28bd</link>
      <guid>https://dev.to/ashthecoder_000005/vibe-manufacturing-is-vibe-coding-with-a-pricier-test-suite-28bd</guid>
      <description>&lt;p&gt;Business Insider just published a fun piece on what people are calling "vibe manufacturing": describing a physical object to a chatbot, getting back a CAD file, and having it fabricated. If you've been anywhere near the "vibe coding" discourse, the shape is instantly familiar.&lt;/p&gt;

&lt;p&gt;The examples, all from &lt;a href="https://www.businessinsider.com/vibe-manufacturing-ai-prototypes-cad-coding-2026-11" rel="noopener noreferrer"&gt;Henry Chandonnet's article&lt;/a&gt;:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;A real estate developer described a steel bracket for his camper by voice, and the chatbot produced a DXF, opened SendCutSend's website, filled in the order form and paid the $156 with his stored payment info. The part arrived ten days later and fit.&lt;/li&gt;
&lt;li&gt;A robotics hobbyist got a sheet-metal robot arm fabricated for about $200, but only after earlier iterations with loose screw holes and "compounding failures" burned through about $2,000 of reorders.&lt;/li&gt;
&lt;li&gt;An IT worker built a storytelling toy for his daughter: a week of modeling instead of months, then he had a second chatbot criticize the first one's design until they reached a consensus, and ordered ten units from JLCPCB for about 500 euros.&lt;/li&gt;
&lt;li&gt;A tinkerer building an answering box for his kids has printed the case around fifteen times and counting. His review: "AI's good but not perfect."&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Worth reading the original. What I want to add is the developer's version, because if you write code, you've seen this exact movie.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdrive.google.com%2Fuc%3Fexport%3Ddownload%26id%3D1fBBIOhOzsXk7HRckyv_L1YEHO7SWApYQ" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdrive.google.com%2Fuc%3Fexport%3Ddownload%26id%3D1fBBIOhOzsXk7HRckyv_L1YEHO7SWApYQ" alt="Flow diagram of the vibe-manufacturing review loop: describe constraints and dimensions, generate CAD with AI, review the dimensions like a diff, test the smallest piece that proves the fit, iterate at days and dollars per loop, then use it. A red no-ship gate routes load-bearing, safety-critical, mains-voltage, food-contact and children's products to qualified engineering review and applicable safety testing first." width="1310" height="849"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The test suite got slower and pricier
&lt;/h2&gt;

&lt;p&gt;With vibe coding, the prompt is the visible ten percent and the gap to production is the work. Vibe manufacturing has the same shape, but the feedback loop changed its physics.&lt;/p&gt;

&lt;p&gt;A unit test runs in milliseconds and costs nothing. A fabricated part takes days and real money, and a wrong dimension doesn't throw an exception. It arrives at your door as plastic or steel. The bracket builder got lucky: he says he was sure it would be a millimeter off, and it wasn't. The robot arm builder paid $2,000 to learn what a failing test feels like when each test run costs $200.&lt;/p&gt;

&lt;p&gt;So the skill that transfers from software isn't prompting. It's the review loop. Everything below is code review and testing habits, translated into atoms.&lt;/p&gt;

&lt;h2&gt;
  
  
  Ask for code, not sculpture
&lt;/h2&gt;

&lt;p&gt;This is the one concrete thing I'd push on if you try this: pay attention to what format the AI gives you back.&lt;/p&gt;

&lt;p&gt;If it gives you a mesh, an STL blob of triangles, you have a statue. You can measure and even edit meshes in mesh software, but the original design history is gone: no named dimensions, no constraints, no record of why that hole is 4mm. A change that should be one number becomes manual sculpting, and when the part comes back wrong the tempting move is to re-roll the prompt and hope.&lt;/p&gt;

&lt;p&gt;If it gives you parametric CAD as code, you have something reviewable. OpenSCAD and CadQuery both describe a part as code: named dimensions, constraints, relationships. OpenSCAD has its own little scripting language; CadQuery is plain Python. You can read the hole diameter before you pay for it. You can change 4mm to 5mm without regenerating the whole part. You can put the design in git and diff revisions, which is a genuinely strange sentence to write about a physical object, and also exactly what you want.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdrive.google.com%2Fuc%3Fexport%3Ddownload%26id%3D1PwPVXQw_CVQ0a-JENyacLZHauqFpNtp9" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdrive.google.com%2Fuc%3Fexport%3Ddownload%26id%3D1PwPVXQw_CVQ0a-JENyacLZHauqFpNtp9" alt="Side-by-side comparison: a mesh or STL file is like a statue. You can measure and edit it in mesh software, but the original design history is gone: no named dimensions to review, no constraint-preserving changes, and a wrong number means manual sculpting or re-rolling the prompt. Parametric CAD as code, like OpenSCAD or CadQuery, is a script: read the hole diameter before you pay, change one number without regenerating the part, and diff revisions in git." width="1310" height="729"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;There's a middle route too: tools like Zoo's Text-to-CAD generate actual B-rep models, real CAD solids rather than meshes. You don't get a readable script that way, but you do get a solid you can open in CAD software, measure properly and modify, which beats a bag of triangles. The Business Insider piece mentions an open-source text-to-CAD project that already tops 23,000 GitHub stars. The direction is clear: the useful tools in this space will be the ones whose output you can inspect and modify, not the ones that hand you a finished-looking blob.&lt;/p&gt;

&lt;p&gt;And note what the toy builder did: he asked a second model to criticize the first model's design. That's a design review. He ran a pull request process on a toy.&lt;/p&gt;

&lt;h2&gt;
  
  
  Print the smallest test first
&lt;/h2&gt;

&lt;p&gt;In software you don't deploy the whole system to find out if one interface is wrong. Same instinct here: before the full part, fabricate the cheapest thing that proves the risky assumption. A hole pattern. A mating surface. A ten-minute print of the one corner where two pieces meet.&lt;/p&gt;

&lt;p&gt;The answering box builder's fifteen prints of the same case sound excessive until you realize that's just the iteration loop, run at home-printer prices instead of fabrication-service prices. Each print was a test run. The bracket that fit perfectly on the first try is the outlier, not the plan.&lt;/p&gt;

&lt;h2&gt;
  
  
  The no-ship list
&lt;/h2&gt;

&lt;p&gt;There's a difference between a hobby prototype and a validated safe design, and every one of these projects sits on the prototype side of that line. The article is honest about it: the more complex the product, the harder this gets, and nobody is vibe manufacturing a car.&lt;/p&gt;

&lt;p&gt;Some things should stay on the no-ship list until they get qualified engineering review and whatever safety and compliance testing applies to them: anything load-bearing, anything safety-critical, anything touching mains voltage, food-contact surfaces, and products for other people's kids. Two chatbots reaching a consensus is a great prototyping trick. It is not an engineering review, and it is not a safety test.&lt;/p&gt;

&lt;p&gt;Also, one detail in the article deserves more attention than it got: the chatbot ordered and paid for the bracket itself, using the developer's stored payment details. An agent that can place fabrication orders with your card needs a written spending boundary before it needs anything else. That part made me more nervous than any of the geometry.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where this actually is today
&lt;/h2&gt;

&lt;p&gt;Read the examples again and a pattern shows up: brackets, jewelry, a brain model, toys, a robot arm the builder calls "edutainment". Personal prototypes and learning projects. The people closest to a product say versions of the same thing: fun to prototype, not ready for a factory.&lt;/p&gt;

&lt;p&gt;That matches where vibe coding was a couple of years ago. The prompt got cheap; the review didn't. So if you try vibe manufacturing, bring the boring skills: read the dimensions, test small, budget for iteration, and keep a human between the AI and anything that has to actually work.&lt;/p&gt;

&lt;p&gt;The machine can speak a part into existence. Somebody still has to check its homework.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>3dprinting</category>
      <category>programming</category>
    </item>
    <item>
      <title>How to Build an AI Agent: From One Prompt to a Production Setup</title>
      <dc:creator>ash the coder</dc:creator>
      <pubDate>Sat, 31 Oct 2026 16:13:30 +0000</pubDate>
      <link>https://dev.to/ashthecoder_000005/how-to-build-an-ai-agent-from-one-prompt-to-a-production-setup-df8</link>
      <guid>https://dev.to/ashthecoder_000005/how-to-build-an-ai-agent-from-one-prompt-to-a-production-setup-df8</guid>
      <description>&lt;p&gt;You have had the moment. You write one prompt, the model nails it, and you think: I could turn this into a product. Then you try, and the gap between a clever prompt and a system you can actually run in production turns out to be almost all of the work.&lt;/p&gt;

&lt;p&gt;This post is that gap, walked end to end.&lt;/p&gt;

&lt;h2&gt;
  
  
  What an agent actually is
&lt;/h2&gt;

&lt;p&gt;Strip away the hype and an agent is three things: a model, instructions, and tools. OpenAI's &lt;a href="https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf" rel="noopener noreferrer"&gt;practical guide to building agents&lt;/a&gt; defines one as a system that independently accomplishes tasks on your behalf, using an LLM to manage the workflow and tools to interact with external systems.&lt;/p&gt;

&lt;p&gt;Anthropic's &lt;a href="https://www.anthropic.com/engineering/building-effective-agents" rel="noopener noreferrer"&gt;Building effective agents&lt;/a&gt; adds the distinction that saves you the most pain:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Workflows&lt;/strong&gt; are LLMs and tools orchestrated through predefined code paths.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Agents&lt;/strong&gt; are LLMs directing their own process and tool usage, deciding their own route through the task.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Their rule of thumb: find the simplest solution possible and only add agentic complexity when it earns its keep. Plenty of things people call agents are really a workflow with one LLM step, and that is fine. It is also cheaper and much easier to debug.&lt;/p&gt;

&lt;p&gt;For this post, one example all the way through: a GitHub issue triage agent. It reads new issues on your repo, applies labels, and drafts a first reply for a maintainer to approve.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 1: Start with one prompt
&lt;/h2&gt;

&lt;p&gt;Before any loop exists, write the system prompt for the job and run it as a single, tool-less call:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;You are a triage assistant for an open source repo.
Read the issue the user pastes in and return JSON with:
- "labels": up to 3 labels from this list only: bug, feature,
  docs, question, duplicate, good-first-issue
- "priority": one of low, medium, high
- "draft_reply": a short, friendly first response
Never invent a label that is not on the list.
If the issue is unclear, say what is missing instead of guessing.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Test it against twenty old issues by hand. If this single call does not label well, an agent will not fix it. It will just fail in more interesting ways.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 2: Give it hands (the loop)
&lt;/h2&gt;

&lt;p&gt;Anthropic describes agents as "typically just LLMs using tools based on environmental feedback in a loop." The whole architecture fits in a paragraph: the model asks for a tool, your code runs it, the result goes back into the conversation, and you repeat until the model says it is done or hits a stopping condition.&lt;/p&gt;

&lt;p&gt;For the triage agent, the tools are small: &lt;code&gt;get_open_issues&lt;/code&gt;, &lt;code&gt;add_labels&lt;/code&gt;, &lt;code&gt;save_draft_reply&lt;/code&gt;. Here is the shape of the loop as a minimal sketch (whatever SDK or model you use, the shape is the same):&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;def&lt;/span&gt; &lt;span class="nf"&gt;run_agent&lt;/span&gt;&lt;span class="p"&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;system_prompt&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="n"&gt;tools&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;max_steps&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;10&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="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;role&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;system&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;content&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;system_prompt&lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;
        &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;role&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;user&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;content&lt;/span&gt;&lt;span class="sh"&gt;"&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="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;_&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="nf"&gt;range&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;max_steps&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
        &lt;span class="n"&gt;response&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="nf"&gt;chat&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;messages&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;tools&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;tool_schemas&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;response&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;tool_calls&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;call&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;response&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;tool_calls&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
                &lt;span class="n"&gt;result&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;tools&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;call&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="n"&gt;call&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;arguments&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
                &lt;span class="n"&gt;messages&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;append&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nf"&gt;tool_message&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;call&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;result&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt;
        &lt;span class="k"&gt;else&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
            &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;response&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt;  &lt;span class="c1"&gt;# the agent decided it is done
&lt;/span&gt;    &lt;span class="k"&gt;raise&lt;/span&gt; &lt;span class="nc"&gt;RuntimeError&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;hit max_steps without finishing&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Keep &lt;code&gt;max_steps&lt;/code&gt;. That line is not decoration. It is the difference between a bug and a bug that bills you.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 3: Check you actually need an agent
&lt;/h2&gt;

&lt;p&gt;Before building anything, OpenAI's guide suggests checking the work against three cases where agents earn their complexity:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Complex decision-making&lt;/strong&gt; with nuanced judgment, like refund approvals.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Rules that got too tangled to maintain&lt;/strong&gt;, like vendor security reviews.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Heavy unstructured data&lt;/strong&gt;, like interpreting documents or free text.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Issue triage qualifies on the first and third: the right label depends on context, and the input is free text written by strangers. If your use case fails all three checks, a deterministic script will beat an agent on cost, speed, and reliability, and it will not hallucinate.&lt;/p&gt;

&lt;h2&gt;
  
  
  Step 4: Guardrails before autonomy
&lt;/h2&gt;

&lt;p&gt;This is the part demo videos skip. Both guides land on the same ideas:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Least-privilege tools.&lt;/strong&gt; Start read-only. Let the agent draft replies, not post them. Write access is something an agent earns after it has behaved.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Human checkpoints.&lt;/strong&gt; Anthropic notes that agents can pause for human feedback at checkpoints or when blocked. For triage, that means drafts go to a queue and a human hits approve.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Stopping conditions.&lt;/strong&gt; Max iterations, a cost ceiling per run, and timeouts. Anthropic is blunt about why: autonomy compounds errors and costs, so they recommend extensive testing in sandboxed environments before agents touch anything real.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Layered defenses.&lt;/strong&gt; OpenAI frames guardrails as layers: input classifiers, tool risk ratings, rules-based protections like blocklists, and output validation, rather than one big filter.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Step 5: The production part nobody posts about
&lt;/h2&gt;

&lt;p&gt;The loop is maybe ten percent of the work. Production is the other ninety:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Evals.&lt;/strong&gt; Build a set of past issues with their correct labels and run the agent against it after every prompt change. Version your prompts like code, because they are.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Traces.&lt;/strong&gt; Log every model call and every tool result. When the agent mislabels something at 3am, the trace is the only way to find out why.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cost and latency budgets.&lt;/strong&gt; Forty cents per triage run is fine for one repo. It is not fine for four hundred.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Deployment.&lt;/strong&gt; Webhook to queue to worker, with results stored for review. Do not run the agent inside the web request; LLM latency will eat your timeouts.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  The honest version
&lt;/h2&gt;

&lt;p&gt;Nobody's first agent is a good agent. The order that works is prompt, then tools, then guardrails, then evals, with autonomy last. Autonomy is the reward for the boring parts being solid.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>python</category>
      <category>tutorial</category>
      <category>llm</category>
    </item>
    <item>
      <title>The Biggest Crypto Heist Ever, and the Three Minutes That Beat the Freeze</title>
      <dc:creator>ash the coder</dc:creator>
      <pubDate>Fri, 30 Oct 2026 08:52:16 +0000</pubDate>
      <link>https://dev.to/ashthecoder_000005/the-biggest-crypto-heist-ever-and-the-three-minutes-that-beat-the-freeze-2988</link>
      <guid>https://dev.to/ashthecoder_000005/the-biggest-crypto-heist-ever-and-the-three-minutes-that-beat-the-freeze-2988</guid>
      <description>&lt;p&gt;It is the biggest crypto heist ever, and the headline everyone is celebrating is the freeze: Tether froze $1.57 billion in stolen USDT, Circle froze over $800 million in USDC, and about half of the $6.3 billion that left Binance is now locked on-chain.&lt;/p&gt;

&lt;p&gt;But inside the same night there is a three-minute transaction that shows exactly where freeze power ends. This is the whole story, as verified as it can be while it is still unfolding.&lt;/p&gt;

&lt;h2&gt;
  
  
  The first alarm
&lt;/h2&gt;

&lt;p&gt;On October 29 at 20:24 UTC, a Binance hot wallet on Ethereum sent about 63 tokens and 167,464 ETH to a single address in one transaction. Minutes later, a second Binance hot wallet sent 28,006 ETH to the same destination, while 18,270 BTC and 613 million USDT on Tron moved to new addresses around the same time. This timeline comes from &lt;a href="https://governance.aave.com/t/binance-hot-wallet-exploit-aaves-response/25954" rel="noopener noreferrer"&gt;Aave's risk team&lt;/a&gt;, who tracked the wallet cluster live, and research firm Kairos Research, cited by &lt;a href="https://cryptoslate.com/trumps-crypto-firm-joins-tether-circle-and-ripple-in-2-6-billion-stablecoin-freeze-as-binance-hack-outflows-near-6-billion/" rel="noopener noreferrer"&gt;CryptoSlate&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Kairos estimates about $5.7 billion left the exchange in total. The attacker used Ethereum's EIP-7702 delegation mechanism to attach executable code to a Binance hot wallet, and a single transaction three minutes later swept the 63 tokens. A second inflow of 444,537 AVAX at 23:04 UTC suggests the compromise was still active more than two hours after it started.&lt;/p&gt;

&lt;p&gt;Binance's first public word was that it was investigating a "potential security issue." Its &lt;a href="https://www.binance.com/en/support/announcement/detail/c5a8729a97ac486d963784d96c5642d7" rel="noopener noreferrer"&gt;official announcement&lt;/a&gt; later put the estimated impact at $6.3 billion, making this the largest crypto hack in history, ahead of the $1.5 billion Bybit hack in 2025.&lt;/p&gt;

&lt;h2&gt;
  
  
  The exits: turning freezable money into unfreezable money
&lt;/h2&gt;

&lt;p&gt;The attacker did not just sit on the funds. According to Etherscan records reviewed by &lt;a href="https://decrypt.co/382194/binance-hack-exit-ramps-defi-exchange-ethereum" rel="noopener noreferrer"&gt;Decrypt&lt;/a&gt; and Aave's risk team, the cluster spread across more than 20 new wallets and started converting, fast:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;On Uniswap, wallets received tens of thousands of ETH in exchange for stablecoins: 20,940 ETH at 20:34 UTC, 3,552 ETH at 20:47, and further batches of roughly 18,085 and 16,498 ETH.&lt;/li&gt;
&lt;li&gt;On Aave, between 20:58 and 21:01 UTC, one wallet supplied 60 million USDT as collateral and borrowed 19,762 ETH against it, worth about $44.6 million. Three minutes, start to finish.&lt;/li&gt;
&lt;li&gt;The cluster also sold governance tokens like UNI, LINK and AAVE for ETH on decentralized exchanges.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The logic is brutal in its simplicity. USDT and USDC are tokens run by companies. Tether and Circle can blacklist an address, and whatever sits in it stops moving. ETH is Ethereum's native asset: no token contract governs it, so there is no issuer that can freeze it on-chain. Every dollar converted before a freeze landed was out of reach.&lt;/p&gt;

&lt;p&gt;The Aave borrow deserves one honest mechanical note. The 60 million USDT he supplied now sits in Aave's shared pool, not in his wallet, and what he holds is a claim on that pool. Aave has no blacklist function, so Tether freezing his address does not automatically freeze that claim. As of Aave's latest update, he has not touched the position since 21:01 UTC, and what happens to the collateral is now a question of liquidation rules and whatever investigators do next. What is already settled: the borrowed ETH is out, no freeze power reaches it, and he has no reason to repay the loan, because repaying would only hand him back a claim on tainted USDT.&lt;/p&gt;

&lt;h2&gt;
  
  
  The freeze
&lt;/h2&gt;

&lt;p&gt;The response was the fastest and largest the industry has ever mounted:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Tether froze nearly $1.57 billion in USDT&lt;/li&gt;
&lt;li&gt;Circle froze over $800 million in USDC (about $823 million)&lt;/li&gt;
&lt;li&gt;Ripple froze roughly $200 million in RLUSD, per Binance co-founder Yi He&lt;/li&gt;
&lt;li&gt;World Liberty Financial restricted $27 million in USD1, and Paxos froze about $8 million in PAX Gold&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;(Figures per &lt;a href="https://www.streetinsider.com/Reuters/Tether%2C+Circle+freeze+%242.38+billion+after+Binance+reports+security+incident/27259548.html" rel="noopener noreferrer"&gt;Reuters&lt;/a&gt; and company statements.) Binance says about $3 billion is frozen on-chain so far, roughly half of what left.&lt;/p&gt;

&lt;p&gt;Aave, whose contracts were never compromised, froze 175 markets across 16 networks on V3 and 117 on V4 between 21:42 and 23:04 UTC to close the conversion route before it could be used again. Existing positions still repay, withdraw and liquidate normally, and the attacker's position remains overcollateralized. As &lt;a href="https://www.cryptotimes.io/2026/10/30/how-aave-was-used-as-a-crypto-mixer-in-binance-hack-attacker-borrowed-44-6m-in-eth-against-60m-usdt/" rel="noopener noreferrer"&gt;contributor Marc Zeller put it&lt;/a&gt;, the protocol will be fine.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who did it, and how they got in
&lt;/h2&gt;

&lt;p&gt;Binance says its initial investigation points to a compromised DevOps engineer's laptop, which let the attacker access a service container and submit unauthorized withdrawal transactions. It found no evidence of a private key leak. Treat this as preliminary: the full technical post-mortem has not been published yet.&lt;/p&gt;

&lt;p&gt;Two analytics firms, TRM Labs and Chainalysis, have attributed the attack with high confidence to a North Korean hacking unit called TraderTraitor, the same cluster of actors linked to the Bybit hack, per the &lt;a href="https://www.bbc.com/news/articles/cwg7xp8kx4v5o" rel="noopener noreferrer"&gt;BBC&lt;/a&gt;. TRM's Ari Redbord called it a "watershed moment" and warned that the attacker still holds an enormous sum, so recovery is far from over.&lt;/p&gt;

&lt;h2&gt;
  
  
  What Binance did
&lt;/h2&gt;

&lt;p&gt;Binance temporarily suspended deposits and withdrawals, revoked the compromised address's authorization, and brought in law enforcement and blockchain intelligence firms. It says it will cover all losses from its SAFU fund and company funds, that user funds are safe, and that it will publish an updated Proof of Reserves once remediation is complete. Withdrawals were expected to resume within about 48 hours.&lt;/p&gt;

&lt;p&gt;Markets flinched: bitcoin fell more than 4% below $78,000, and the governance tokens the attacker was dumping fell harder (AAVE and LINK down about 9%, UNI about 8%, per Aave's risk team at 00:55 UTC). Stablecoins held their pegs.&lt;/p&gt;

&lt;h2&gt;
  
  
  The honest scoreboard
&lt;/h2&gt;

&lt;p&gt;It would be easy to read this night as "freezes are useless." The numbers say otherwise: about half of $6.3 billion was frozen within hours, an unprecedented coordinated response across multiple issuers and chains. The Aave conversion was $44.6 million of the total.&lt;/p&gt;

&lt;p&gt;But the freeze has a speed limit, and it is measured in minutes. The attacker needed three. DeFi runs 24/7, has no compliance button, and will lend anyone ETH against stablecoin collateral at 3am on a holiday. In a crisis, every permissionless lending pool and DEX is a potential conversion endpoint from freezable money to unfreezable money.&lt;/p&gt;

&lt;p&gt;If you build anything that takes stablecoins as collateral, deposits or payment, you just watched your threat model play out in production. Your protocol is not a bystander in the next incident. It is infrastructure.&lt;/p&gt;

&lt;p&gt;The freeze saved billions. Three minutes on Aave showed where it ends. Both are true, and building like only the first one is true is the mistake this heist leaves behind.&lt;/p&gt;

&lt;h2&gt;
  
  
  What we still do not know
&lt;/h2&gt;

&lt;p&gt;How the engineer's laptop was compromised. How much of the remaining billions is recoverable. Whether the attacker can extract the Aave collateral. And the full post-mortem, which Binance says comes after the investigation. If you are reading this weeks later, check those updates before quoting any number here.&lt;/p&gt;

</description>
      <category>security</category>
      <category>web3</category>
      <category>cryptocurrency</category>
      <category>blockchain</category>
    </item>
    <item>
      <title>Meta spent $35 billion on servers and data centers and its earnings call never said 'Llama'</title>
      <dc:creator>ash the coder</dc:creator>
      <pubDate>Thu, 29 Oct 2026 08:26:10 +0000</pubDate>
      <link>https://dev.to/ashthecoder_000005/meta-spent-35-billion-on-servers-and-data-centers-and-its-earnings-call-never-said-llama-dad</link>
      <guid>https://dev.to/ashthecoder_000005/meta-spent-35-billion-on-servers-and-data-centers-and-its-earnings-call-never-said-llama-dad</guid>
      <description>&lt;p&gt;Two things stand out from Meta's Q3 earnings call: the company reported almost $35 billion of capital expenditure in a single quarter servers, data centers and network infrastructure  and the word "Llama" did not come up once in the entire call.&lt;/p&gt;

&lt;p&gt;A year ago, every Meta AI conversation was about open weights and Llama. This call was all about one word: Muse. Here is what Meta actually announced, in plain English.&lt;/p&gt;

&lt;h2&gt;
  
  
  Meet Muse
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F507jqarxnvgg9lhf3z5b.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F507jqarxnvgg9lhf3z5b.jpg" alt="Meta's Muse character on its companion device (image: Meta, via Creative Bloq)" width="800" height="450"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Muse is Meta's personal AI agent the product form of what Zuckerberg calls personal superintelligence. It launched about six weeks before the call, and according to Meta:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;More than 10 million people in the US and Canada have set up a Muse agent&lt;/li&gt;
&lt;li&gt;It has been #1 in the App Store almost every day since launch&lt;/li&gt;
&lt;li&gt;About 25% of people who set it up use it on any given day&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Those are Meta's own numbers, so take them with a pinch of salt. But even with a discount, 10 million setups in six weeks is a fast start.&lt;/p&gt;

&lt;p&gt;Behind the agent is a whole new model family, all Muse-branded:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Muse Wonder&lt;/strong&gt; - their first "frontier" model, which Zuckerberg called state-of-the-art "across many dimensions" (his claim, not an independent benchmark)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Muse Spark 1.2 and 1.3&lt;/strong&gt;, &lt;strong&gt;Muse Code&lt;/strong&gt;, and &lt;strong&gt;Muse Voice&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;Two open-weights models: &lt;strong&gt;Muse Glimmer&lt;/strong&gt; and &lt;strong&gt;Muse Spark Open&lt;/strong&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;So open weights is not dead at Meta. It just got a new name. Bigger models are in training, and smaller "product models" with Wonder-level intelligence are coming soon, per the call.&lt;/p&gt;

&lt;h2&gt;
  
  
  The North Star quote
&lt;/h2&gt;

&lt;p&gt;Bernstein analyst Mark Shmulik asked Zuckerberg on the call whether Meta is reorienting the whole company around Muse. Zuckerberg's answer is worth reading:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"It was more than a year ago that I ... wrote the letter about personal superintelligence as the North Star for the work that we're doing here."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;He added that Meta can build the best experience by owning the end-to-end technology itself the models and the agents and closed with the line that tells you how he thinks about all of this: "as long as we continue to do that, then over time the business takes care of itself."&lt;/p&gt;

&lt;p&gt;Translation: Meta wants to own the whole stack the model, the agent, and the device. Muse starts rolling out on Meta's smart glasses next month. The glasses are the delivery mechanism, the agent is the product, and the model is the moat.&lt;/p&gt;

&lt;h2&gt;
  
  
  The money, in plain English
&lt;/h2&gt;

&lt;p&gt;This is where it gets wild. From Meta's official Q3 numbers:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Metric&lt;/th&gt;
&lt;th&gt;Q3 2026&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Revenue&lt;/td&gt;
&lt;td&gt;$64.6B, up 26% year over year&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Costs and expenses&lt;/td&gt;
&lt;td&gt;$51.6B, up 68%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Operating income&lt;/td&gt;
&lt;td&gt;Down 37%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Free cash flow&lt;/td&gt;
&lt;td&gt;Negative $2.1B&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Capex this quarter&lt;/td&gt;
&lt;td&gt;$34.9B (servers, data centers, network)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;2026 capex forecast&lt;/td&gt;
&lt;td&gt;$135-145B&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The core ad business is very healthy. The spending is what is wild. Meta is forecasting $135-145 billion of capital expenditure this year while its quarterly free cash flow went negative.&lt;/p&gt;

&lt;p&gt;One fairness note: the expense jump includes a one-off $9.89 billion legal charge a settlement with 52 US state attorneys general over teen mental health and social media which is also why Meta raised its full-year expense forecast to $175-179 billion.&lt;/p&gt;

&lt;p&gt;Meta says it expects Muse to get more than 10x cheaper to run within a year, and to be highly profitable over multiple years if compute efficiency keeps improving. That is a plan, not a guarantee - and the call said the quiet part out loud: the strength of the ad business is what funds all of it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why this matters if you build things
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;The agent is the new interface.&lt;/strong&gt; Meta's apps reach 3.64 billion people a day, and Zuckerberg explicitly called WhatsApp "an increasingly important platform for agents". If Meta's bet works, a lot of user interactions on those platforms will go through Muse - worth thinking about if you build anything that touches Instagram or WhatsApp audiences.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Meta wants to sell the pickaxes too.&lt;/strong&gt; It is building a Meta Enterprise Platform to serve inference demand from its own models - a new name to watch in the model-API space.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Open weights continues, rebranded.&lt;/strong&gt; Llama was not mentioned once on the call; open-weights Muse models were. If you build on Llama today, the branding and roadmap are clearly shifting - keep an eye on what happens to the Llama line.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The scale of the bet is the story.&lt;/strong&gt; Whatever you think of Meta, a company going free-cash-flow negative to build out data centers tells you where big tech thinks this is all going.&lt;/li&gt;
&lt;/ol&gt;

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

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Third-Quarter-2026-Results/default.aspx" rel="noopener noreferrer"&gt;Meta Q3 2026 results (official press release)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://s21.q4cdn.com/399680738/files/content_files/META-Q3-2026-Earnings-Call-Transcript.pdf" rel="noopener noreferrer"&gt;Meta Q3 2026 earnings call transcript (official PDF)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://finance.yahoo.com/technology/ai/articles/meta-north-star-mark-zuckerberg-003009701.html" rel="noopener noreferrer"&gt;Yahoo Finance write-up on the call&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>news</category>
      <category>llm</category>
    </item>
    <item>
      <title>A viral "how X ranks your posts" thread, checked against X's actual open-source code</title>
      <dc:creator>ash the coder</dc:creator>
      <pubDate>Tue, 27 Oct 2026 11:34:24 +0000</pubDate>
      <link>https://dev.to/ashthecoder_000005/a-viral-how-x-ranks-your-posts-thread-checked-against-xs-actual-open-source-code-1edi</link>
      <guid>https://dev.to/ashthecoder_000005/a-viral-how-x-ranks-your-posts-thread-checked-against-xs-actual-open-source-code-1edi</guid>
      <description>&lt;p&gt;A thread has been going around promising a breakdown of "the new X algorithm update": exact multipliers for every reaction. A reply is worth 7.6 likes! A copied link is 30! A report is -355!&lt;/p&gt;

&lt;p&gt;X publishes the code that ranks the For You feed at &lt;a href="https://github.com/xai-org/x-algorithm" rel="noopener noreferrer"&gt;xai-org/x-algorithm&lt;/a&gt;, and it is genuinely readable. This post checks each claim in the thread against it.&lt;/p&gt;

&lt;p&gt;One caveat up front: everything below is the default values in the published code, pinned to a specific commit (&lt;a href="https://github.com/xai-org/x-algorithm/tree/2f252142b9284e692ac72851768c8072d045f0d5" rel="noopener noreferrer"&gt;2f25214&lt;/a&gt;, checked 27 Oct 2026). X's production config can override any of it. Treat this as the published design, not a live config dump.&lt;/p&gt;

&lt;p&gt;The short version: the numbers are real. The obvious reading of them is wrong. And one claim is out of date.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the numbers come from
&lt;/h2&gt;

&lt;p&gt;The ranking weights live in &lt;code&gt;home-mixer/params/param.rs&lt;/code&gt;. Current defaults:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Action&lt;/th&gt;
&lt;th&gt;Weight in the code&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Copy link&lt;/td&gt;
&lt;td&gt;+20.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Reply from a mutual, on an original post&lt;/td&gt;
&lt;td&gt;+5.0 with a +15.0 boost&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Reply&lt;/td&gt;
&lt;td&gt;+5.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Quote&lt;/td&gt;
&lt;td&gt;+5.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Share via DM&lt;/td&gt;
&lt;td&gt;+5.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Follow author&lt;/td&gt;
&lt;td&gt;+4.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Share&lt;/td&gt;
&lt;td&gt;+2.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Repost&lt;/td&gt;
&lt;td&gt;+1.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Like&lt;/td&gt;
&lt;td&gt;+0.66&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Block author&lt;/td&gt;
&lt;td&gt;-31.2&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;"Not interested"&lt;/td&gt;
&lt;td&gt;-52.272&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Mute author&lt;/td&gt;
&lt;td&gt;-58.8&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Report&lt;/td&gt;
&lt;td&gt;-234.0&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F9l91970vv67rntyh69d8.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F9l91970vv67rntyh69d8.png" alt="Positive ranking weights in X's open-source code" width="799" height="560"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Divide everything by the like weight (0.66) and you get the thread's numbers: 20 ÷ 0.66 ≈ 30, 5 ÷ 0.66 ≈ 7.6, -234 ÷ 0.66 ≈ -355. The thread re-normalized the weights so a like equals 1. So yes: real numbers, real code.&lt;/p&gt;

&lt;p&gt;Bookmarks are the funny one. The thread says a bookmark counts as 0. The truth is stronger: there is no bookmark score in the code at all. The model doesn't even predict one.&lt;/p&gt;

&lt;h2&gt;
  
  
  The part that's easy to get wrong
&lt;/h2&gt;

&lt;p&gt;Reading this table as "one reply equals 7.6 likes" is exactly the mistake the code's own comments warn against. From &lt;code&gt;param.rs&lt;/code&gt;, verbatim:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"it'd be incorrect to see that a report has 468 times higher weight than a like and conclude that e.g. '1 report cancels out 468 likes'. The weights are a multiple on your own predicted probability of Liking, Reporting, etc."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The formula is not a count of what happened. It is:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;score = Σ (weight × P(action))&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;where P(action) is the probability that &lt;em&gt;you, the specific viewer&lt;/em&gt;, will take that action, predicted by a transformer model (Phoenix) from your engagement history. The same post gets a different score for every single person who might see it. There is no universal "this post is worth N points."&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpmwoe9nzv9u5xew0s5va.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpmwoe9nzv9u5xew0s5va.png" alt="How the For You feed scores a post" width="800" height="556"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Two consequences the comments spell out:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Mass-reporting doesn't sink a post globally.&lt;/strong&gt; "Recommendations are personalized, so reports from bad actors will primarily affect recommendations for users who are similar to the bad actors." Also, a report is over 1,000x rarer than a like, which is why its weight is so large - a rare action needs a big weight to matter in the sum at all.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Engagement from a group chat link does nothing.&lt;/strong&gt; "Directly navigating to a post (i.e., coordinating via groupchat) has no ranking impact." Only engagement from a post that was actually served in the Home feed counts.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fkj3ke57i9eq3jlyldlpd.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fkj3ke57i9eq3jlyldlpd.png" alt="Negative ranking weights in X's open-source code" width="800" height="573"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The small-account boost: the thread nailed the numbers
&lt;/h2&gt;

&lt;p&gt;The most practically useful claims in the thread check out exactly, from the cold-start params:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;ColdStartFollowerCap&lt;/code&gt; = 50,000 followers&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;ColdStartMaxPostAgeSecs&lt;/code&gt; = 7,200 (2 hours)&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;ColdStartImpressionThreshold&lt;/code&gt; = 200 impressions&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;ColdStartSlotMin&lt;/code&gt; / &lt;code&gt;ColdStartSlotMax&lt;/code&gt; = 15-16&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F83hd295djaf5nh1jb1ru.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F83hd295djaf5nh1jb1ru.png" alt="X's cold-start boost parameters" width="800" height="577"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Read this as eligibility, not a guarantee: under these defaults, an original post from an account under 50K followers, under two hours old, with under 200 impressions, is eligible to be lifted toward slot 15-16 in a feed. Whether production runs exactly these defaults, and whether any given post gets the lift, is not something the code can tell us.&lt;/p&gt;

&lt;p&gt;Also confirmed: posts from accounts you don't follow get multiplied by 0.75 (&lt;code&gt;OonWeightFactor&lt;/code&gt; in &lt;code&gt;vm-ranker/params.rs&lt;/code&gt;), and replies and reposts from accounts you do follow get rescaled too (&lt;code&gt;EnableOonRescoreForInNetworkRepliesRetweets&lt;/code&gt; = true).&lt;/p&gt;

&lt;h2&gt;
  
  
  What the thread gets stale
&lt;/h2&gt;

&lt;p&gt;The thread says that when your posts compete for one feed, the second keeps 62.5% of its score and the third keeps 43.75%. Current defaults: &lt;code&gt;AuthorDiversityDecay&lt;/code&gt; = 0.5 with &lt;code&gt;AuthorDiversityFloor&lt;/code&gt; = 0.25. So the second post keeps 50%, and later posts bottom out at 25% - not 43.75%. Those viral numbers may match an older commit; they are not what's published now.&lt;/p&gt;

&lt;p&gt;And "after 48 hours X stops putting your post in feeds" overstates one constant. What exists is &lt;code&gt;SimclustersMaxCandidateAgeHours&lt;/code&gt; = 48: out-of-network discovery candidates older than 48 hours get dropped. That is not a global kill switch. Posts from accounts people follow live in a different pipeline (thunder), with no 48-hour cutoff published.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I'd take from this (judgment, not hacks)
&lt;/h2&gt;

&lt;p&gt;None of this is a cheat code, because the score is personal and the weights multiply probabilities. With that said, the shape of the defaults suggests a few boring things:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Write things people want to reply to, quote, or copy the link of. Those actions carry the biggest positive weights. (Copy link is the largest single weight at +20; the +15 is additive on top of a reply's +5, and neither is a promise - a weight times a near-zero probability is still near zero.)&lt;/li&gt;
&lt;li&gt;Make originals easy for your mutuals to reply to. The +15 boost applies when someone you mutually follow replies to your original post. You can't manufacture that, but questions and genuine invitations make it likelier.&lt;/li&gt;
&lt;li&gt;If the defaults hold, a post's first two hours are when the small-account boost can apply - so posting when you're around to talk back is sensible.&lt;/li&gt;
&lt;li&gt;Don't fire off five posts in a row. After your first post in someone's feed, each extra one from you is multiplied by 0.5, down to a floor of 0.25 - bursts shrink fast, though never to zero.&lt;/li&gt;
&lt;li&gt;"Not interested", mutes and reports carry the biggest negative weights. But their effect still depends on the viewer's predicted probabilities - these weights are large, not absolute.&lt;/li&gt;
&lt;/ul&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://github.com/xai-org/x-algorithm/blob/2f252142b9284e692ac72851768c8072d045f0d5/README.md" rel="noopener noreferrer"&gt;xai-org/x-algorithm README&lt;/a&gt; - how the feed is assembled and scored&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://github.com/xai-org/x-algorithm/blob/2f252142b9284e692ac72851768c8072d045f0d5/home-mixer/params/param.rs" rel="noopener noreferrer"&gt;home-mixer/params/param.rs&lt;/a&gt; - every weight, plus the comments warning against misreading them&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://github.com/xai-org/x-algorithm/blob/2f252142b9284e692ac72851768c8072d045f0d5/vm-ranker/params.rs" rel="noopener noreferrer"&gt;vm-ranker/params.rs&lt;/a&gt; - author diversity and out-of-network adjustments&lt;/li&gt;
&lt;li&gt;&lt;a href="https://x.com/shannholmberg/status/2114780241810067903" rel="noopener noreferrer"&gt;The thread this fact-checks&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;All values checked against commit &lt;code&gt;2f25214&lt;/code&gt; on 27 Oct 2026. The file changes often - if you're reading this months from now, open &lt;code&gt;param.rs&lt;/code&gt; again before quoting any of it.&lt;/p&gt;

</description>
      <category>opensource</category>
      <category>algorithms</category>
      <category>webdev</category>
      <category>career</category>
    </item>
    <item>
      <title>The Bug That Sat in PHP for a Decade, Until AI Cracked It in Weeks</title>
      <dc:creator>ash the coder</dc:creator>
      <pubDate>Fri, 23 Oct 2026 08:35:56 +0000</pubDate>
      <link>https://dev.to/ashthecoder_000005/the-bug-that-sat-in-php-for-a-decade-until-ai-cracked-it-in-weeks-3dn5</link>
      <guid>https://dev.to/ashthecoder_000005/the-bug-that-sat-in-php-for-a-decade-until-ai-cracked-it-in-weeks-3dn5</guid>
      <description>&lt;p&gt;On Wednesday, Finland took its government websites offline. The Dutch parliament went dark. Edinburgh council's site disappeared. Parts of Wikipedia stopped loading. Even a few BBC programme pages went down.&lt;/p&gt;

&lt;p&gt;The reason: two bugs in PHP, the language that quietly runs most of the web. The vulnerable code had been sitting there for more than a decade. Then AI-assisted researchers turned it into working exploits within weeks, and everything kicked off.&lt;/p&gt;

&lt;p&gt;This is the story of PHP2Shell, and I think it's the most interesting security story of the year, because the good guys and the bad guys are about to use the exact same trick.&lt;/p&gt;

&lt;h2&gt;
  
  
  First, what actually broke
&lt;/h2&gt;

&lt;p&gt;PHP powers roughly 70% of websites, according to &lt;a href="https://www.wiz.io/blog/php2shell-php-rce" rel="noopener noreferrer"&gt;Wiz's estimate&lt;/a&gt;. WordPress, Drupal, Joomla, MediaWiki (the software behind Wikipedia) and a mountain of shared hosting and control panels all sit on top of it. If you run WordPress or WooCommerce, which is the stack I build on, your server runs PHP.&lt;/p&gt;

&lt;p&gt;On October 21, the PHP team patched two critical vulnerabilities, &lt;a href="https://www.cisa.gov/news-events/alerts/2026/10/21/critical-vulnerabilities-php-require-immediate-patching" rel="noopener noreferrer"&gt;CVE-2026-109618 and CVE-2026-109619&lt;/a&gt;, both rated CVSS 9.2. Researchers named them PHP2Shell, which tells you the punchline: unauthenticated remote code execution. An attacker doesn't need a login, and they don't need a bug in your code. The bugs are in PHP itself, so &lt;a href="https://www.php2shell.com/" rel="noopener noreferrer"&gt;even a "Hello World" app on a vulnerable version can be exposed&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;The technical idea is elegant in a nasty way. When PHP hits a fatal error, like running out of memory mid-request, it can leave corrupted internal state behind. The worker process survives, and the next request handled by that same worker can inherit the stale state. Chain it right and you get &lt;a href="https://www.wiz.io/blog/php2shell-php-rce" rel="noopener noreferrer"&gt;use-after-free primitives that turn into full code execution&lt;/a&gt;, demonstrated on default WordPress, Drupal and Joomla installs.&lt;/p&gt;

&lt;p&gt;One of the bugs is worse in a quieter way. It lets an attacker read the PHP worker's memory remotely, including passwords, tokens and API keys from other people's requests, and the stolen data comes back inside a &lt;code&gt;PHPSESSID&lt;/code&gt; cookie. Wiz compared it to Heartbleed, and the comparison is fair. No login, no code execution needed. Ask nicely and the server hands you its secrets.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fe2lja5sxejr8zthsfn9m.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fe2lja5sxejr8zthsfn9m.png" alt="Wiz demonstrating PHP2Shell remote code execution against WordPress, Drupal and Joomla" width="800" height="462"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Image: &lt;a href="https://www.wiz.io/blog/php2shell-php-rce" rel="noopener noreferrer"&gt;Wiz Research&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The part that reads like fiction
&lt;/h2&gt;

&lt;p&gt;The vulnerable code survived more than a decade of humans reading it. There's a reason for that, and it isn't that nobody cared. According to &lt;a href="https://www.tomshardware.com/tech-industry/cyber-security/openai-says-gpt-5-6-sol-helped-uncover-critical-php-flaws-that-could-affect-70-percent-of-websites-ai-assisted-researchers-turned-two-year-old-bug-that-stumped-human-developers-into-working-exploits-in-two-weeks" rel="noopener noreferrer"&gt;Tom's Hardware&lt;/a&gt;, the PHP Foundation had known something was wrong with the underlying code since 2024, but couldn't identify an exploitation method or a fix. The chain was too complex to demonstrate: too many steps, too much fiddly memory layout work.&lt;/p&gt;

&lt;p&gt;Then two teams, working independently, pointed frontier AI models at it.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://openai.com/index/coordinated-disclosure-of-critical-vulnerabilities" rel="noopener noreferrer"&gt;Calif&lt;/a&gt;, a security testing firm in OpenAI's "Patch the Planet" program, began auditing PHP with GPT-5.6 Sol on July 30, at the PHP Foundation's request. Researcher Ryan A. identified the exploitable vulnerability on August 11, with OpenAI researchers finding it the same day, per Tom's Hardware's reporting of OpenAI's account.&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://www.wiz.io/blog/php2shell-php-rce" rel="noopener noreferrer"&gt;Wiz&lt;/a&gt; found related bugs separately using its Atlas agent, Google's Gemini 4 Argon and GPT-5.6 Sol. Its own timeline: first discovery August 19, working RCE against Drupal the next day, WordPress the day after, everything reported to PHP by August 24.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Wiz's claim stuck with me: "a single researcher completed in a week what previously would have taken a world-class team months of work." The bugs weren't new. What changed is that exploiting them stopped being impractical.&lt;/p&gt;

&lt;h2&gt;
  
  
  Then it leaked, and attackers moved fast
&lt;/h2&gt;

&lt;p&gt;The plan was a calm, coordinated disclosure. Patches and technical details were scheduled for October 22, with defenders briefed in advance. That's not what happened.&lt;/p&gt;

&lt;p&gt;On October 20, &lt;a href="https://www.bbc.com/news/articles/cmly0lg3w5x2o" rel="noopener noreferrer"&gt;details of the flaws leaked on social media&lt;/a&gt;, and the release was pulled forward a day. Separately, and more worrying: &lt;a href="https://arstechnica.com/security/2026/10/hackers-got-a-four-day-head-start-on-epic-php2shell-vulnerability/" rel="noopener noreferrer"&gt;Amazon said it identified hackers attempting to exploit the bug four days before public disclosure&lt;/a&gt;, and described the flaws as "now being broadly exploited in the wild."&lt;/p&gt;

&lt;p&gt;Two honesty notes here. First, those are exploitation attempts reported by Amazon via the Financial Times and Ars Technica; CISA said it &lt;a href="https://www.cisa.gov/news-events/alerts/2026/10/21/critical-vulnerabilities-php-require-immediate-patching" rel="noopener noreferrer"&gt;had not observed exploitation as of its October 21 advisory&lt;/a&gt;, and I've seen no confirmed breach attributed to these bugs yet. Attempts are not the same as a successful compromise. Second, the "AI found it first" claim needs care. It's &lt;a href="https://openai.com/index/coordinated-disclosure-of-critical-vulnerabilities" rel="noopener noreferrer"&gt;OpenAI's account&lt;/a&gt;, and it covers discovery: the private window was real, about two months from the August discoveries to the October patch. But the public window was essentially zero. Amazon says exploitation attempts were already happening four days before disclosure, so by the time the world knew, attackers were already moving.&lt;/p&gt;

&lt;p&gt;Either way, Wednesday looked dramatic. The &lt;a href="https://www.nsa.gov/Press-Room/Press-Releases-Statements/Press-Release-View/Article/4630238/nsa-advisory-critical-patches-recommended-for-php-vulnerabilities/" rel="noopener noreferrer"&gt;NSA, CISA and FBI&lt;/a&gt; warned organizations to patch immediately or take systems offline, with the NSA saying the vulnerabilities "often provide a direct path to the server's file system or database." The UK's NCSC, CERT-EU, and agencies in Australia and Japan issued similar advice. Japan's Prime Minister personally posted on X urging people to check whether their sites run PHP, &lt;a href="https://www.bbc.com/news/articles/cmly0lg3w5x2o" rel="noopener noreferrer"&gt;per the BBC&lt;/a&gt;. When a head of government is doing patch comms, it's a real one.&lt;/p&gt;

&lt;p&gt;And it's not just governments. People are noticing on their own feeds:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Flh3.googleusercontent.com%2Fd%2F1udAd6b06_BoE4LTDeHeFF2aZEuizmjEc" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Flh3.googleusercontent.com%2Fd%2F1udAd6b06_BoE4LTDeHeFF2aZEuizmjEc" alt="A post on X reading " width="1179" height="1148"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;A post on X this morning. The poster's name and photo are cropped out, and the screenshot ends mid-sentence; the full post links to &lt;a href="https://www.php.net/archive/2026.php" rel="noopener noreferrer"&gt;php.net's advisory archive&lt;/a&gt;. Source: &lt;a href="https://x.com/Tieshaaaah/status/2113519539916869862" rel="noopener noreferrer"&gt;@Tieshaaaah on X&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The honest bit: AI did not simply save the day
&lt;/h2&gt;

&lt;p&gt;It would be easy to read this as an AI triumph story. It isn't, quite.&lt;/p&gt;

&lt;p&gt;On OpenAI's account, defenders got there first on discovery. OpenAI calls that lead time a &lt;a href="https://openai.com/index/coordinated-disclosure-of-critical-vulnerabilities" rel="noopener noreferrer"&gt;"Defender's Window"&lt;/a&gt; and used it to coordinate a genuinely impressive response: WAF rules ready at Cloudflare, AWS, Akamai and Fastly before public disclosure, government agencies engaged, and $1 billion committed in subsidized security tooling for under-resourced defenders.&lt;/p&gt;

&lt;p&gt;But Wiz is blunt about the flip side: &lt;a href="https://www.wiz.io/blog/php2shell-php-rce" rel="noopener noreferrer"&gt;"With AI this is no longer a barrier. LLMs can help adapt an exploit to different PHP versions, configurations and memory layouts."&lt;/a&gt; The same capability that let Calif and Wiz build these exploits is the capability that makes the bugs dangerous now. For a decade, these flaws were protected by the sheer difficulty of exploiting them. That protection just expired, for defenders and attackers alike.&lt;/p&gt;

&lt;p&gt;Also, one caveat on scale: the "70%" figure is Wiz's estimate of PHP's share of websites, and the &lt;a href="https://www.bbc.com/news/articles/cmly0lg3w5x2o" rel="noopener noreferrer"&gt;BBC noted it couldn't verify it&lt;/a&gt;. Usage share also isn't the same as exploitable share, and nobody has published a solid count of how many servers are still unpatched. Even so, the vulnerable footprint is enormous by any measure.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I'd actually do (the boring checklist)
&lt;/h2&gt;

&lt;p&gt;If you run anything PHP, and if you self-host WordPress or WooCommerce that's you:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Patch the runtime, not the app.&lt;/strong&gt; Updating WordPress does nothing here. Install PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35, then restart Apache, PHP-FPM and any long-running PHP workers, per the &lt;a href="https://www.php2shell.com/" rel="noopener noreferrer"&gt;official guidance&lt;/a&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;If you can't patch, take it offline.&lt;/strong&gt; That's not me being dramatic: CISA says organizations that can't update, or that run PHP 8.1 or earlier (which get no fix), &lt;a href="https://www.cisa.gov/news-events/alerts/2026/10/21/critical-vulnerabilities-php-require-immediate-patching" rel="noopener noreferrer"&gt;should take PHP-based systems offline until they can&lt;/a&gt;. The NSA and NCSC said the same.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A WAF is a seatbelt, not a fix.&lt;/strong&gt; CISA says one "may provide some protection until systems are updated." &lt;a href="https://aws.amazon.com/blogs/security/using-aws-waf-to-apply-mitigations-for-php2shell-rce-vulnerabilities-cve-2026-109618-and-cve-2026-109619/" rel="noopener noreferrer"&gt;AWS&lt;/a&gt;, Cloudflare and Akamai shipped managed rules. Use them while you patch, not instead of patching.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Find every place PHP hides.&lt;/strong&gt; It's under hosting panels, control panels, monitoring tools and appliances that never mention it. Sophos told the BBC the hard part isn't patching, it's &lt;a href="https://www.bbc.com/news/articles/cmly0lg3w5x2o" rel="noopener noreferrer"&gt;finding every instance where PHP is running&lt;/a&gt;. OpenAI released a &lt;a href="https://openai.com/index/coordinated-disclosure-of-critical-vulnerabilities" rel="noopener noreferrer"&gt;version checker as a Codex skill&lt;/a&gt; if you want help inventorying.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Check whether someone got there first.&lt;/strong&gt; Amazon reports active exploitation attempts. Wiz's blog lists the tells: unusual multipart requests followed by memory-limit errors, repeated HTTP 500s, weirdly large or malformed &lt;code&gt;PHPSESSID&lt;/code&gt; cookies, and unexplained PHP worker restarts.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  The bigger question
&lt;/h2&gt;

&lt;p&gt;Wiz framed PHP2Shell as a Heartbleed or Log4Shell moment for the AI era, and I think that's right, but for a different reason than most coverage says. Those bugs taught us the internet runs on shared infrastructure nobody audits. This one teaches us that "too hard to exploit" was never a security property. It was a labour cost, and AI just made that labour cheap.&lt;/p&gt;

&lt;p&gt;OpenAI says it &lt;a href="https://openai.com/index/coordinated-disclosure-of-critical-vulnerabilities" rel="noopener noreferrer"&gt;expects similar findings to increase sharply in the coming months&lt;/a&gt;. I believe them. The open question is who finds the next one first, and how wide the Defender's Window stays when both sides have the same models.&lt;/p&gt;

&lt;p&gt;What do you think: does the window hold, or was this the widest it will ever be? And if you run PHP anywhere, how painful was your Wednesday?&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Sources: &lt;a href="https://www.wiz.io/blog/php2shell-php-rce" rel="noopener noreferrer"&gt;Wiz Research&lt;/a&gt;, &lt;a href="https://openai.com/index/coordinated-disclosure-of-critical-vulnerabilities" rel="noopener noreferrer"&gt;OpenAI&lt;/a&gt;, &lt;a href="https://www.php2shell.com/" rel="noopener noreferrer"&gt;php2shell.com&lt;/a&gt;, &lt;a href="https://www.cisa.gov/news-events/alerts/2026/10/21/critical-vulnerabilities-php-require-immediate-patching" rel="noopener noreferrer"&gt;CISA&lt;/a&gt;, &lt;a href="https://www.nsa.gov/Press-Room/Press-Releases-Statements/Press-Release-View/Article/4630238/nsa-advisory-critical-patches-recommended-for-php-vulnerabilities/" rel="noopener noreferrer"&gt;NSA&lt;/a&gt;, &lt;a href="https://www.bbc.com/news/articles/cmly0lg3w5x2o" rel="noopener noreferrer"&gt;BBC&lt;/a&gt;, &lt;a href="https://arstechnica.com/security/2026/10/hackers-got-a-four-day-head-start-on-epic-php2shell-vulnerability/" rel="noopener noreferrer"&gt;Ars Technica / FT&lt;/a&gt;, &lt;a href="https://techcrunch.com/2026/10/21/php-armageddon-openai-googles-wiz-find-ancient-holes-panicking-the-internet/" rel="noopener noreferrer"&gt;TechCrunch&lt;/a&gt;, &lt;a href="https://www.tomshardware.com/tech-industry/cyber-security/openai-says-gpt-5-6-sol-helped-uncover-critical-php-flaws-that-could-affect-70-percent-of-websites-ai-assisted-researchers-turned-two-year-old-bug-that-stumped-human-developers-into-working-exploits-in-two-weeks" rel="noopener noreferrer"&gt;Tom's Hardware&lt;/a&gt;, &lt;a href="https://aws.amazon.com/blogs/security/using-aws-waf-to-apply-mitigations-for-php2shell-rce-vulnerabilities-cve-2026-109618-and-cve-2026-109619/" rel="noopener noreferrer"&gt;AWS Security Blog&lt;/a&gt;. Wiz's announcement post: &lt;a href="https://infosec.exchange/@wiz/117479149813368302" rel="noopener noreferrer"&gt;https://infosec.exchange/@wiz/117479149813368302&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

</description>
      <category>security</category>
      <category>php</category>
      <category>ai</category>
      <category>webdev</category>
    </item>
  </channel>
</rss>
