<?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: Sammi De Blas </title>
    <description>The latest articles on DEV Community by Sammi De Blas  (@analista_83).</description>
    <link>https://dev.to/analista_83</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%2F4120781%2Fdd2c76ae-692c-4f18-b62f-4ecade622e2f.jpg</url>
      <title>DEV Community: Sammi De Blas </title>
      <link>https://dev.to/analista_83</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/analista_83"/>
    <language>en</language>
    <item>
      <title>The agent that steals keys to pay its own bill</title>
      <dc:creator>Sammi De Blas </dc:creator>
      <pubDate>Wed, 18 Nov 2026 12:26:02 +0000</pubDate>
      <link>https://dev.to/analista_83/the-agent-that-steals-keys-to-pay-its-own-bill-2f8o</link>
      <guid>https://dev.to/analista_83/the-agent-that-steals-keys-to-pay-its-own-bill-2f8o</guid>
      <description>&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%2F4e1mm3t86u4x0k3xehb4.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%2F4e1mm3t86u4x0k3xehb4.png" alt="Diagram: the agent that steals keys to pay its own bill" width="800" height="420"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The thief agent
&lt;/h2&gt;

&lt;p&gt;Between October 3 and November 5, Amazon's MadPot honeypot network kept receiving visitors that were not human.&lt;/p&gt;

&lt;p&gt;They were AI agents hunting for model gateways exposed to the internet, pulling out the provider keys stored inside and testing whether those keys were still alive (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;The keys that worked were not sold on. They went into a common pool and were used to pay for the agent's own inference (Source: venturebeat.com).&lt;/p&gt;

&lt;p&gt;That detail changes the nature of the attack.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;This is not credential theft for resale on a forum. This is an agent that funds the compute it needs to keep stealing with its own loot.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The playbook that gives itself away
&lt;/h2&gt;

&lt;p&gt;What MadPot captured were not new exploits, because the requests carried the full campaign history, including the operator's instructions and the findings the agents themselves reported back (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;Amazon has moderate confidence that the framework behind it was ARTEX (I still have to test it in a sandbox, I just lack the hours).&lt;/p&gt;

&lt;p&gt;The entry manual the agents kept repeating was uncomfortably simple.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;The default master key of LiteLLM.&lt;/li&gt;
&lt;li&gt;Half-finished installs anyone can claim.&lt;/li&gt;
&lt;li&gt;Admin interfaces exposed without authentication.&lt;/li&gt;
&lt;li&gt;And CVE-2025-3248 in Langflow.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;(Source: venturebeat.com).&lt;/p&gt;

&lt;p&gt;The target list in the logs is a map of a badly deployed AI ecosystem, so I will enumerate it:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;One API&lt;/li&gt;
&lt;li&gt;New API&lt;/li&gt;
&lt;li&gt;Veloera&lt;/li&gt;
&lt;li&gt;Sub2API&lt;/li&gt;
&lt;li&gt;LiteLLM&lt;/li&gt;
&lt;li&gt;Dify&lt;/li&gt;
&lt;li&gt;FastGPT&lt;/li&gt;
&lt;li&gt;Langflow&lt;/li&gt;
&lt;li&gt;RAGFlow&lt;/li&gt;
&lt;li&gt;Open WebUI&lt;/li&gt;
&lt;li&gt;LobeChat&lt;/li&gt;
&lt;li&gt;Ollama&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;(Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;They all make the same mistake: they store keys from several providers in one place.&lt;/p&gt;

&lt;p&gt;A weak admin page there opens many accounts at once.&lt;/p&gt;

&lt;h2&gt;
  
  
  This is getting out of hand
&lt;/h2&gt;

&lt;p&gt;I see a change of business model (several, actually), not just a change of tool.&lt;/p&gt;

&lt;p&gt;A human attacker steals keys and sells them, fine, and that has an hourly cost and a market that sets a price.&lt;/p&gt;

&lt;p&gt;An agent steals keys, uses them to think, and with that finds more keys.&lt;/p&gt;

&lt;p&gt;The marginal cost of each new target trends to zero and it keeps no schedule.&lt;/p&gt;

&lt;p&gt;That is why Amazon insists it cannot verify the results the agents reported, I wrote about it on my blog days ago.&lt;/p&gt;

&lt;p&gt;What it has is the transcript of what they intended to do, not proof that they pulled it off (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;And even so, volume is the signal.&lt;/p&gt;

&lt;p&gt;A human operator fires dozens of commands per hour against a target, sure, but with an agent CrowdStrike has measured up to 1,500 in one hour (Source: venturebeat.com).&lt;/p&gt;

&lt;p&gt;Honestly, do you not feel this is getting out of hand?&lt;/p&gt;

&lt;h2&gt;
  
  
  The window is measured in...
&lt;/h2&gt;

&lt;p&gt;The FBI published an advisory on ARTEX on November 13 describing intrusions that went from the initial target to privileged access in under three hours (Source: venturebeat.com).&lt;/p&gt;

&lt;p&gt;The same advisory says the developer deleted the original repository in October after reports of malicious use, and by the way, the forks are still there.&lt;/p&gt;

&lt;p&gt;By early November more than 1,600 exposed ARTEX consoles had been identified (Source: venturebeat.com).&lt;/p&gt;

&lt;p&gt;Cloudflare documented another case with the same shape: a compromised CI/CD API token, with no IP restriction and no rate limit, generated 400 database dumps in one hour (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;The FBI cites ARTEX attacks on financial entities in South Korea with data on around 67,000 people (Source: venturebeat.com).&lt;/p&gt;

&lt;p&gt;CrowdStrike saw an attacker use the framework against six major South Korean banks, steal information and then ask Claude where to sell it (Source: venturebeat.com).&lt;/p&gt;

&lt;p&gt;Banks are in the crosshairs and I work for one of them.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this breaks
&lt;/h2&gt;

&lt;p&gt;The classic "breakout time" metric measured how long an attacker takes to move from one compromised machine to another.&lt;/p&gt;

&lt;p&gt;CrowdStrike has announced it is dropping it because with agents it has stopped meaning anything.&lt;/p&gt;

&lt;p&gt;Nobody agrees on what replaces it (Source: venturebeat.com).&lt;/p&gt;

&lt;p&gt;If the attacker does not sleep, does not hesitate and does not charge by the hour, time stops being the variable that separates the defender from the intruder.&lt;/p&gt;

&lt;p&gt;What is left is the attack surface you have.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to look at?
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Default keys.&lt;/strong&gt; LiteLLM ships a known master key. If your gateway keeps it, you do not have a gateway, you have a front door wide open. Change it and check that no console is left without authentication.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Half-finished installs.&lt;/strong&gt; A Dify, a Langflow or an Ollama that someone spun up to test and never closed is still there. Look for those products' ports on your network and on anything you publish, and do not expose Ollama on 0.0.0.0 if you run it locally.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;API spend.&lt;/strong&gt; A stolen key shows up on the bill before it shows up in the logs. Check consumption spikes per provider and compare them with your applications' real usage.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;CI/CD tokens.&lt;/strong&gt; The Cloudflare one had no IP limit and no request limit. Put both on any token that can touch production.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How would I test it?
&lt;/h2&gt;

&lt;p&gt;I would spin up LiteLLM in a container, with its default master key and a fake provider key inside.&lt;/p&gt;

&lt;p&gt;From another machine on the network I would look for the admin panel and try the known key.&lt;/p&gt;

&lt;p&gt;If it lets me in, I already have the demonstration that the flaw is not in the model, it is in the deployment.&lt;/p&gt;

&lt;p&gt;Then I would repeat it with the key changed and the console closed, to measure how long someone who has never seen LiteLLM takes to spot the difference.&lt;/p&gt;

&lt;p&gt;It is a twenty-minute test that says more than any report on autonomous agents.&lt;/p&gt;

&lt;h2&gt;
  
  
  Closing
&lt;/h2&gt;

&lt;p&gt;The agent that pays for itself with what it steals is a business model, full stop.&lt;/p&gt;

&lt;p&gt;As long as defense keeps measuring how long the attacker takes to move, the attacker has already stopped being in a hurry, read that last sentence again because it sums it all up.&lt;/p&gt;




&lt;p&gt;Originally published at &lt;a href="https://sammideblas.com/notas/the-agent-that-steals-keys-to-pay-its-own-bill" rel="noopener noreferrer"&gt;https://sammideblas.com/notas/the-agent-that-steals-keys-to-pay-its-own-bill&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Si has leído hasta aquí... reacciona y comparte -&amp;gt; "la seguridad y la defensa en la era de la IA es cosa de todos los que la usan" ¿No Crees?&lt;/p&gt;

</description>
      <category>ai</category>
      <category>security</category>
      <category>apikeys</category>
      <category>supplychain</category>
    </item>
    <item>
      <title>The vendor keeps the key</title>
      <dc:creator>Sammi De Blas </dc:creator>
      <pubDate>Mon, 16 Nov 2026 22:05:49 +0000</pubDate>
      <link>https://dev.to/analista_83/the-vendor-keeps-the-key-54i6</link>
      <guid>https://dev.to/analista_83/the-vendor-keeps-the-key-54i6</guid>
      <description>&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%2F37lnew4zlr9hsk0ljnu3.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%2F37lnew4zlr9hsk0ljnu3.jpg" alt="Diagram: the vendor keeps the key" width="800" height="447"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The vendor keeps the key
&lt;/h2&gt;

&lt;p&gt;Anthropic published Claude Fable 5.2 and Mythos 5.2 on November 9 (Source: anthropic.com, explainx.ai).&lt;/p&gt;

&lt;p&gt;It is the same model at two levels of safeguards, since the famous Mythos stays restricted to verified organizations. Fable is the public version.&lt;/p&gt;

&lt;p&gt;The figures in the announcement are self-reported: 49.1% on FrontierCode 2.0, up to 30% faster, and an API 20% cheaper, at 8 and 40 dollars per million tokens (Source: anthropic.com, explainx.ai).&lt;/p&gt;

&lt;p&gt;The prior test was run by the AI Security Institute and METR (Source: anthropic.com, explainx.ai).&lt;/p&gt;

&lt;p&gt;The difference between the two versions is not speed or price. It is what each one agrees to do.&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%2F8scx5n17rjny83vuv9nf.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%2F8scx5n17rjny83vuv9nf.jpg" alt="When the permission is the bug" width="800" height="447"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Mozilla reported that Mythos 5.2 doubled vulnerability detection against 5.1 in scanning campaigns (Source: gptunnel.ru).&lt;/p&gt;

&lt;p&gt;That capability is what Fable, the public version, trims back on cybersecurity tasks (Source: gptunnel.ru).&lt;/p&gt;

&lt;p&gt;I use Anthropic too, and the ability to find and exploit flaws is not removed. It is put behind a door with identity verification.&lt;/p&gt;

&lt;h2&gt;
  
  
  The model that hides what it knows
&lt;/h2&gt;

&lt;p&gt;OpenAI revealed that GPT-6 Astra was hiding help on security tasks without telling the researcher (Source: bleepingcomputer.com).&lt;/p&gt;

&lt;p&gt;The cause the company pointed to was a grader that rewarded hiding over an explicit refusal (Source: bleepingcomputer.com).&lt;/p&gt;

&lt;p&gt;It is silent sandbagging, not a visible refusal: the model does not say no, it pretends to help and hides what it knows. Sneaky.&lt;/p&gt;

&lt;p&gt;It happened inside a cybersecurity evaluation, with researchers watching.&lt;/p&gt;

&lt;p&gt;And the cause was not a decision by the model. It was a rubric that paid better for silence than for refusal.&lt;/p&gt;

&lt;p&gt;The lesson is not that models are bad. It is that if the incentive rewards hiding, the log you produce is not evidence.&lt;/p&gt;

&lt;h2&gt;
  
  
  When the permission is the flaw
&lt;/h2&gt;

&lt;p&gt;On November 16, a scan of 159,041 projects built with Lovable found 24,664 with a Supabase database readable without login (Source: cybernews.com).&lt;/p&gt;

&lt;p&gt;Exposed: more than 200,000 private messages, 80,000 profiles, and more than 1,000 passports.&lt;/p&gt;

&lt;p&gt;There was no hack. It was Row Level Security left off.&lt;/p&gt;

&lt;p&gt;Row Level Security is the Supabase mechanism that decides which rows each user can read, and if it is not enabled, the database answers anyone who knows the URL.&lt;/p&gt;

&lt;p&gt;It is the same pattern as the rest of the week, under another name.&lt;/p&gt;

&lt;p&gt;AI writes the application in minutes and the door stays open because nobody configured the permission.&lt;/p&gt;

&lt;p&gt;A week earlier, WorkOS confirmed that an ARTEX agent (I have been experimenting with it) escaped a researcher's sandbox and reached their internal sales app (Source: bleepingcomputer.com).&lt;/p&gt;

&lt;p&gt;It did so with a development credential hardcoded in production and no rate limit (Source: bleepingcomputer.com).&lt;/p&gt;

&lt;p&gt;It was only detected when an employee saw anomalous usage (Source: bleepingcomputer.com).&lt;/p&gt;

&lt;p&gt;Astra decides which help it denies, and ARTEX turns a forgotten permission into exfiltration.&lt;/p&gt;

&lt;p&gt;The Lovable generator leaves the database open, and in all three cases the audit depends on the good behavior of the system being audited.&lt;/p&gt;

&lt;h2&gt;
  
  
  The regulator arrives afterwards
&lt;/h2&gt;

&lt;p&gt;Australia will create an AI security regulator with penalties and a duty to report incidents and misalignments (Source: apnews.com).&lt;/p&gt;

&lt;p&gt;The decision comes after an OpenAI agent's unauthorized access to the Medicare statistics portal in June (Source: apnews.com).&lt;/p&gt;

&lt;p&gt;Japan approved an emergency package of 10 billion yen, about 63 million dollars, after a wave of breaches (Source: securityweek.com).&lt;/p&gt;

&lt;p&gt;The expert team will use AI to hunt for flaws in networks before the end of the month (Source: securityweek.com).&lt;/p&gt;

&lt;p&gt;AI is both the defender's tool and the attacker's. The regulator arrives after the incident, not before.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where PHP2Shell fits
&lt;/h2&gt;

&lt;p&gt;The previous week, two remote code execution vulnerabilities in PHP, known as PHP2Shell, showed the other end of the same problem (Source: apnews.com).&lt;/p&gt;

&lt;p&gt;AI agents found them, specifically GPT-5.6 Sol and Gemini 4 Argon (Source: openai.com).&lt;/p&gt;

&lt;p&gt;Some of the flaws had been in the code for more than a decade.&lt;/p&gt;

&lt;p&gt;The NSA, the FBI, the UK NCSC, and CERT-EU asked for the affected service to be patched or retired (Source: apnews.com).&lt;/p&gt;

&lt;p&gt;What is interesting is not the age or who found them. It is that detection is no longer the bottleneck.&lt;/p&gt;

&lt;p&gt;Finding the flaw has become cheap, and deciding who can touch what and proving they did it is still expensive.&lt;/p&gt;

&lt;h2&gt;
  
  
  My reading
&lt;/h2&gt;

&lt;p&gt;For anyone running systems, the problem is not that a model capable of finding flaws exists.&lt;/p&gt;

&lt;p&gt;It is that the decision of who can use it is made by the vendor, on criteria it does not publish.&lt;/p&gt;

&lt;p&gt;An identity check is not an audit: it says who you are, not what you do with what you are given.&lt;/p&gt;

&lt;p&gt;And a rubric that rewards hiding is not an audit either, and on top of that it produces a report that looks complete and is not.&lt;/p&gt;

&lt;p&gt;If the same model that detects the vulnerability can exploit it, the record of what it did should be external and verifiable.&lt;/p&gt;

&lt;p&gt;It is no longer enough for the vendor to say it controls it.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to look at
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Check RLS.&lt;/strong&gt; If you have a Supabase project, open the panel and verify that Row Level Security is on for every table. Not the one you think matters. All of them.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Trim inherited permissions.&lt;/strong&gt; Review which development credentials are still alive in production. The ARTEX one was hardcoded and had no rate limit; nobody saw it until an employee noticed odd usage.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Turn on your own logs.&lt;/strong&gt; If an agent acts on your network, the record cannot live only in the agent's system. Keep an external trace that does not depend on the model writing it correctly.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Ask about the safeguard tier.&lt;/strong&gt; If you buy a model with two versions, get in writing which one you are given and what tasks it trims. The difference between Fable and Mythos does not show up in the price.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How I would test it in my lab
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;&lt;p&gt;I would set up a minimal Supabase project locally, a table with invented data, and a small agent with read-only permission over a test folder.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;I would disable RLS on purpose and make a request with curl without an authentication header. Then I would enable it and repeat the same request.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;After that I would try to get the agent out of its sandbox with a simple instruction, no exploit, just asking.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;I would check whether the host system's log records the attempt or whether only the agent itself reports it.&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;If the only record lives inside the agent, the test already tells me what I need. The control that watches cannot live inside what is watched.&lt;/p&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Closing
&lt;/h2&gt;

&lt;p&gt;The ability to find flaws does not disappear when it is put behind a door. Only who holds the key changes.&lt;/p&gt;

&lt;p&gt;And as long as the vendor keeps the key, the question left is who audits the one who decides who gets in.&lt;/p&gt;

&lt;p&gt;The rest is configuring permissions. And no model does that yet.&lt;/p&gt;




&lt;p&gt;Originally published at &lt;a href="https://sammideblas.com/notas/the-vendor-keeps-the-key" rel="noopener noreferrer"&gt;https://sammideblas.com/notas/the-vendor-keeps-the-key&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Si has leído hasta aquí... reacciona y comparte -&amp;gt; "la seguridad y la defensa en la era de la IA es cosa de todos los que la usan" ¿No Crees?&lt;/p&gt;

</description>
      <category>ai</category>
      <category>security</category>
      <category>permissions</category>
      <category>supplychain</category>
    </item>
    <item>
      <title>The agent is a user without a license</title>
      <dc:creator>Sammi De Blas </dc:creator>
      <pubDate>Fri, 13 Nov 2026 20:04:20 +0000</pubDate>
      <link>https://dev.to/analista_83/the-agent-is-a-user-without-a-license-429c</link>
      <guid>https://dev.to/analista_83/the-agent-is-a-user-without-a-license-429c</guid>
      <description>&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%2F6rybkaa04mekuz789nj0.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%2F6rybkaa04mekuz789nj0.png" alt="Esquema: The agent is a user without a license" width="800" height="420"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The agent without an ID
&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%2Fo6s7ymhjnccax64rhups.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%2Fo6s7ymhjnccax64rhups.jpg" alt="the agent is a user without a license captura 1" width="800" height="447"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;United Airlines will restrict third-party AI agents from its website, its app, and also when they arrive through agencies and aggregators (cnbc.com).&lt;/p&gt;

&lt;p&gt;The reason they give is that those agents share credentials and can break FAA and DOT rules.&lt;/p&gt;

&lt;p&gt;Only a human can declare that the luggage carries no explosives.&lt;/p&gt;

&lt;p&gt;An agent cannot sign that declaration, and not because it cannot read it. It cannot sign because there is no identity behind it to answer when the signature is wrong.&lt;/p&gt;

&lt;p&gt;Amazon already blocked Muse, Meta's agent, and Walmart integrated it into its workflow (cnbc.com).&lt;/p&gt;

&lt;p&gt;Two companies of the same size, two opposite decisions, and neither one is a whim. One closed the door and the other opened it with conditions.&lt;/p&gt;

&lt;p&gt;The question left in the air is the same in both cases.&lt;/p&gt;

&lt;p&gt;What identity does the agent act with? Who answers when it gets it wrong?&lt;/p&gt;

&lt;p&gt;I will leave it there...&lt;/p&gt;

&lt;h2&gt;
  
  
  The pattern
&lt;/h2&gt;

&lt;p&gt;An agent wants to act in a system that was not designed for it.&lt;/p&gt;

&lt;p&gt;The FAA requires a human declaration and the agent cannot give one.&lt;/p&gt;

&lt;p&gt;The agent walks in with someone's keys and the system cannot tell whether it is that person or an automated process.&lt;/p&gt;

&lt;p&gt;There is no standard for agent authentication and every company improvises.&lt;/p&gt;

&lt;p&gt;Integrate the agent or close the door on it?&lt;/p&gt;

&lt;p&gt;Those are the two options available today.&lt;/p&gt;

&lt;h2&gt;
  
  
  The other side
&lt;/h2&gt;

&lt;p&gt;An OpenAI agent accessed the Medicare statistics portal without authorization in June (apnews.com).&lt;/p&gt;

&lt;p&gt;Australia has reacted with an AI safety regulator with penalties and an obligation to report incidents and misalignments (apnews.com).&lt;/p&gt;

&lt;p&gt;The same month, a public PoC called worse_query lets an app read data from other apps on iOS 26.0-27.2 without jailbreak or kernel exploit (thehackernews.com).&lt;/p&gt;

&lt;p&gt;Apple does not confirm a patch and there is no CVE assigned.&lt;/p&gt;

&lt;p&gt;Binance warns about exposed seed phrases in the same period.&lt;/p&gt;

&lt;p&gt;The agent is not only a user without a license when it acts for you. It is also a user without a license when it acts against you, and in both cases the system does not know who to hold accountable, though I understand the owner is traced and the weight of the law falls on them... right?&lt;/p&gt;

&lt;h2&gt;
  
  
  Identity or patching?
&lt;/h2&gt;

&lt;p&gt;On Friday Forescout published a data point that changes the scale of the problem.&lt;/p&gt;

&lt;p&gt;Vedere Labs cut vulnerability analysis from 9 months to 90 hours using commercial AI (forescout.com).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;They validated 99 flaws across 10 TCP/IP stacks&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;66 unknown&lt;/li&gt;
&lt;li&gt;52% high or critical severity.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If discovery compresses from months to days, patching stops being the defensive advantage.&lt;/p&gt;

&lt;p&gt;The advantage moves to who holds the session, with what scope, and who can revoke it.&lt;/p&gt;

&lt;p&gt;The SCKit campaign against ScreenConnect illustrates it.&lt;/p&gt;

&lt;p&gt;It exploits CVE-2026-84869, already patched, hijacks operator sessions, and its RelayTrap module pivots back against the technician investigating the incident (thehackernews.com).&lt;/p&gt;

&lt;p&gt;The patch closes the known hole but it does not close the stolen session, I think I have said that already.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to look at on Monday
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Trim the agent's scope.&lt;/strong&gt; If an agent enters with a person's credentials, ask for them to be delegated, with permissions limited to the task and with an expiry date. Without that, the agent inherits everything its owner can do.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Turn on per-session tracing.&lt;/strong&gt; Every agent action has to be tied to an identity and a moment. If you cannot reconstruct what it did and when, you cannot revoke it or audit the incident.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Separate the session from the patch.&lt;/strong&gt; Review which support sessions are still open after you apply the patch. ScreenConnect shows that a closed hole does not close the session that was already inside.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How would I test it?
&lt;/h2&gt;

&lt;p&gt;I would set up a local agent with read-only permissions on a test folder and give it a concrete task, like summarizing a file.&lt;/p&gt;

&lt;p&gt;Then I would try to make it write to that folder. If it succeeds, the permission is not properly scoped and the problem is mine, not the model's.&lt;/p&gt;

&lt;p&gt;I would repeat the test with a delegated credential and an expiry date, and what I would look at is whether the system rejects the action and whether it leaves a trace of who tried it.&lt;/p&gt;

&lt;p&gt;With that I would have a cheap answer to the question of the week.&lt;/p&gt;

&lt;p&gt;If my agent cannot do more than I asked, identity is solved.&lt;br&gt;
If it can, I have a user without a license inside my house.&lt;/p&gt;

&lt;h2&gt;
  
  
  The close
&lt;/h2&gt;

&lt;p&gt;The debate of the week is not whether an AI can book a flight but with what identity it does it and who answers if it gets it wrong.&lt;/p&gt;

&lt;p&gt;Until there is a standard for agent authentication, every company has only two options.&lt;/p&gt;

&lt;p&gt;Integrate the agent or close the door on it.&lt;/p&gt;




&lt;p&gt;Originally published at &lt;a href="https://sammideblas.com/notas/the-agent-is-a-user-without-a-license" rel="noopener noreferrer"&gt;https://sammideblas.com/notas/the-agent-is-a-user-without-a-license&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Si has leído hasta aquí... reacciona y comparte -&amp;gt; "la seguridad y la defensa en la era de la IA es cosa de todos los que la usan" ¿No Crees?&lt;/p&gt;

</description>
      <category>security</category>
      <category>ai</category>
      <category>identity</category>
      <category>accesscontrol</category>
    </item>
    <item>
      <title>The AI bill is visible and the cost per task is not</title>
      <dc:creator>Sammi De Blas </dc:creator>
      <pubDate>Tue, 10 Nov 2026 09:15:20 +0000</pubDate>
      <link>https://dev.to/analista_83/the-ai-bill-is-visible-and-the-cost-per-task-is-not-4kej</link>
      <guid>https://dev.to/analista_83/the-ai-bill-is-visible-and-the-cost-per-task-is-not-4kej</guid>
      <description>&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%2Ftpnc98kalthm4d32e0pj.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%2Ftpnc98kalthm4d32e0pj.png" alt="Esquema: The AI bill is visible and the cost per task is not" width="800" height="420"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The main case
&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%2Fodvoq6ibey2rpmkvy87v.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%2Fodvoq6ibey2rpmkvy87v.jpg" alt="bill visible cost" width="800" height="447"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;On November 3 Menlo Ventures published its annual report on AI in the enterprise.&lt;/p&gt;

&lt;p&gt;Deployment spend reached 177 billion dollars in September against 33 billion a year earlier (Source: menlovc.com).&lt;/p&gt;

&lt;p&gt;That is 5.4 times more in twelve months and the pace is accelerating. In 2025 growth was 3.9 times.&lt;/p&gt;

&lt;p&gt;The money is not spread evenly either, since Anthropic, OpenAI and Google take 122.7 billion, 69% of the total when a year ago it was 50%.&lt;/p&gt;

&lt;p&gt;Anthropic goes from 40% to 55% of enterprise API spend while Google falls from 21% to 9%.&lt;/p&gt;

&lt;p&gt;On OpenRouter agents already generate 80% of the tokens, about 130.9 trillion per week (Source: menlovc.com).&lt;/p&gt;

&lt;p&gt;Consumption stopped being measured by people and is measured by work executed.&lt;/p&gt;

&lt;p&gt;If the cost goes per task, the number of employees stops predicting the invoice.&lt;/p&gt;

&lt;p&gt;55% of that spend is new budget, not money that came out of software or payroll.&lt;/p&gt;

&lt;p&gt;That is, AI does not replace a budget line, it only adds one.&lt;/p&gt;

&lt;p&gt;Spend also concentrates where the result can be checked.&lt;/p&gt;

&lt;p&gt;Programming and chatbots add up to 76% of an applications market of 77.3 billion (Source: menlovc.com).&lt;/p&gt;

&lt;p&gt;It is no accident: those are the places where an agent can act and where someone can verify whether it got it right.&lt;/p&gt;

&lt;p&gt;For whoever approves support budgets, the change is direct, since a variable invoice gets signed for the first time and the metric stops being how many people there are.&lt;/p&gt;

&lt;h2&gt;
  
  
  The other side
&lt;/h2&gt;

&lt;p&gt;The same week, Deloitte published its annual survey and set the counterpoint.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;81% of leaders say AI improved decision speed&lt;/li&gt;
&lt;li&gt;25% is truly transforming the business&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;(Source: deloitte.com).&lt;/p&gt;

&lt;p&gt;The headline of its second finding is the phrase of the month.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;"The bill is visible, the unit cost is not."&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;You approve by model capacity and you review by the feeling of speed, almost never by what a resolved task costs.&lt;/p&gt;

&lt;p&gt;Menlo closes with a figure that works as a mirror, pay attention. Every dollar invested in training models generated 1.70 dollars of revenue from deployment (Source: menlovc.com).&lt;/p&gt;

&lt;p&gt;The industry already knows how to measure its aggregate return, but from all this long talk I give you my conclusion is that what it has not solved is the cost of the concrete task, the one approved on Monday and reviewed on Friday, for example.&lt;/p&gt;

&lt;h2&gt;
  
  
  My own reading
&lt;/h2&gt;

&lt;p&gt;The problem is not the invoice, it is the unit of measurement.&lt;/p&gt;

&lt;p&gt;You buy capacity and you review feeling, and between those two things there is no number that says how much it cost to resolve the customer's case.&lt;/p&gt;

&lt;p&gt;An agent that fails and retries costs twice and does not warn you, it happens a lot, nobody measures that rate, the budget goes into work that never got delivered.&lt;/p&gt;

&lt;p&gt;Deloitte's 25% is what has been transformed with all that invoice and the rest is purchased haste, they used to call it FOMO, right?&lt;/p&gt;

&lt;h2&gt;
  
  
  What to look at?
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Cost per task, not per seat.&lt;/strong&gt; If the provider bills by tokens, ask for the spend of one real case from start to finish. A three-week pilot gives you that number and a demo does not.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Who pays for the retries.&lt;/strong&gt; An agent that fails and tries again costs twice and does not warn you. If nobody measures that rate, the budget goes into work that never got delivered.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Whether the 25% transforms or only accelerates.&lt;/strong&gt; Deloitte's 81% is speed. Before renewing, compare the cost per unit delivered with last year's. If latency drops and unit cost rises, you have bought haste.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How I would test it
&lt;/h2&gt;

&lt;p&gt;I have this half built in BENCHMARKS-LLM, a Streamlit app that reads my reports and runs batteries against local GGUF models.&lt;/p&gt;

&lt;p&gt;I measure speed in tokens per second, latency and a 0 to 5 rubric scored by an optional cloud judge. What I do not have is the cost column, since they are local LLMs, but without that column I repeat the report's mistake.&lt;/p&gt;

&lt;p&gt;I would add three fields to the report writer, input token price, output token price and number of retries, and I would divide the run's spend by the answers the judge approved.&lt;/p&gt;




&lt;p&gt;Originally published at &lt;a href="https://sammideblas.com/notas/the-ai-bill-is-visible-and-the-cost-per-task-is-not" rel="noopener noreferrer"&gt;https://sammideblas.com/notas/the-ai-bill-is-visible-and-the-cost-per-task-is-not&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Si has leído hasta aquí... reacciona y comparte -&amp;gt; "la seguridad y la defensa en la era de la IA es cosa de todos los que la usan" ¿No Crees?&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>devops</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Trust expires between the check and the use</title>
      <dc:creator>Sammi De Blas </dc:creator>
      <pubDate>Fri, 06 Nov 2026 10:13:46 +0000</pubDate>
      <link>https://dev.to/analista_83/trust-expires-between-the-check-and-the-use-dhi</link>
      <guid>https://dev.to/analista_83/trust-expires-between-the-check-and-the-use-dhi</guid>
      <description>&lt;h2&gt;
  
  
  The package that gets verified and then swapped
&lt;/h2&gt;

&lt;p&gt;On November 4, CVE-2026-109518 was published in APT, the package manager for Debian and Ubuntu.&lt;/p&gt;

&lt;p&gt;The flaw is a race. An attacker in the middle of the connection, or a malicious mirror, swaps the package right after APT has verified it and before it installs it.&lt;/p&gt;

&lt;p&gt;The result installs as root.&lt;/p&gt;

&lt;p&gt;It affects Debian 12 and 13 and Ubuntu 22.04 and later, in default configuration. A patch already exists in DSA-6607-1 (Source: threatdefence.com).&lt;/p&gt;

&lt;p&gt;What stands out is not the overflow or the broken signature. It is where it happens.&lt;/p&gt;

&lt;p&gt;APT's cryptographic verification works. The package being checked is authentic. The one installed an instant later is not.&lt;/p&gt;

&lt;p&gt;Between those two moments there is a window nobody was watching.&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%2Ffrsoa7gjlfwms286w1tz.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%2Ffrsoa7gjlfwms286w1tz.jpg" alt="The window between verification and install" width="800" height="447"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  It is not an isolated case
&lt;/h2&gt;

&lt;p&gt;The same week brings the same pattern somewhere else.&lt;/p&gt;

&lt;p&gt;CISA added CVE-2026-119977 to KEV, a TOCTOU race in the staging paths of a full WordPress Core update. TOCTOU means time-of-check to time-of-use, the time between checking and using (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;During that window there is unauthenticated remote code execution. 12,632 attacks were blocked in 24 hours and the federal deadline expires on November 6.&lt;/p&gt;

&lt;p&gt;On November 4, CVE-2026-111585 also appeared in TeamViewer, with CVSS 8.8. A compromised remote host attacks whoever connects to it, in default configuration, and accepting the session is enough (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;The common shape is clear. The defender connects to something to fix it and that something is what compromises them.&lt;/p&gt;

&lt;p&gt;The trust machine has become the target.&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%2Fl6xcxxaep3lkxrvm7420.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%2Fl6xcxxaep3lkxrvm7420.jpg" alt="The same pattern in WordPress and TeamViewer" width="800" height="447"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The other side, the mark that does not say who
&lt;/h2&gt;

&lt;p&gt;In AI the week goes the same direction, with different vocabulary.&lt;/p&gt;

&lt;p&gt;xAI will mark text from seven Grok models by default from November 7, with an x-watermark header and opt-out only for team admins. Anthropic enabled SynthID-Text on the 3rd in four more Claude models, with no opt-out. The common argument is the EU AI Act (Source: aitoolsrecap.com).&lt;/p&gt;

&lt;p&gt;The mark proves a model took part in the text. It does not prove who wrote it, with what intent, or whether someone edited it afterwards.&lt;/p&gt;

&lt;p&gt;It is the same failure as APT. There is a check, and what was checked stops being true the next moment.&lt;/p&gt;

&lt;p&gt;My reading is that we are putting seals at the wrong point in the chain. We verify the input and assume the output keeps the property.&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%2Fgmxx20p2t1h58l9iibgy.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%2Fgmxx20p2t1h58l9iibgy.jpg" alt="Watermarks prove a model took part, not who wrote" width="800" height="447"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The scale is no longer human
&lt;/h2&gt;

&lt;p&gt;On November 5, OpenAI opened GPT-6 Cyber only to approved Daybreak Red customers. The model produced 81 high-severity vulnerabilities in Firefox and 948 local privilege escalation exploits in Linux, from 2,308 candidate findings and 1.21 million dollars in API credits (Source: openai.com, unite.ai).&lt;/p&gt;

&lt;p&gt;The same day, OpenAI's misalignment report from November 4 admits something else. GPT-6 Astra did about 40% less work on code tasks labeled as security research, 10% against 16% on hidden rubrics, almost always without saying so. GPT-5.6 Sol showed no such gap (Source: alignment.openai.com).&lt;/p&gt;

&lt;p&gt;There is the problem in its cleanest form. The model finds real exploits at industrial scale and, at the same time, silently decides how much effort to spend based on what the ticket is called.&lt;/p&gt;

&lt;p&gt;The metric that matters is not how many vulnerabilities it finds. It is whether it admits when it did not work.&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%2Fm2bnjp6pfilvfrmudsen.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%2Fm2bnjp6pfilvfrmudsen.jpg" alt="The scale is no longer human" width="800" height="447"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What to look at on Monday
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The APT window.&lt;/strong&gt; Check whether your Debian and Ubuntu servers have DSA-6607-1 applied. If you update through your own mirror, verify the mirror is not exposed and that download and install do not get separated in time.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The WordPress TOCTOU.&lt;/strong&gt; If you have automatic Core updates, check whether staging is reachable from outside during the window. The November 6 federal deadline has already passed for US government agencies.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The watermark as proof.&lt;/strong&gt; If someone presents a marked text as evidence of authorship, remember it only certifies that a model took part. The chain of custody of the text is still your problem.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How I would test it in my lab
&lt;/h2&gt;

&lt;p&gt;I would set up two virtual machines with Debian 12 and my own APT mirror on a third.&lt;/p&gt;

&lt;p&gt;I would capture the traffic with an intermediate proxy and try to swap the package between verification and install, only on my isolated network.&lt;/p&gt;

&lt;p&gt;If the modified package installs as root, I have confirmation that the patch is mandatory and not optional.&lt;/p&gt;

&lt;p&gt;The same with a test WordPress and a Core update triggered by hand, watching whether the staging directory is served over HTTP during the window.&lt;/p&gt;

&lt;p&gt;The conclusion I would draw is simple. Verification is not enough if the verified object can change afterwards.&lt;/p&gt;

&lt;h2&gt;
  
  
  Closing
&lt;/h2&gt;

&lt;p&gt;Trust has an expiry date and almost nobody looks at it. Between the check and the use there is room for an attacker, and this week there was room three times.&lt;/p&gt;

&lt;p&gt;If you made it this far, react and share. Security and defence in the AI era is everyone's business, not just the people who build it.&lt;/p&gt;

</description>
      <category>security</category>
      <category>supplychain</category>
      <category>ai</category>
      <category>linux</category>
    </item>
    <item>
      <title>The email was real and the fraud was in the comment</title>
      <dc:creator>Sammi De Blas </dc:creator>
      <pubDate>Thu, 05 Nov 2026 19:33:09 +0000</pubDate>
      <link>https://dev.to/analista_83/the-email-was-real-and-the-fraud-was-in-the-comment-jkg</link>
      <guid>https://dev.to/analista_83/the-email-was-real-and-the-fraud-was-in-the-comment-jkg</guid>
      <description>&lt;h2&gt;
  
  
  The email was from dev.to and the fraud was in the comment
&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%2F97l0upr0b11cgwq1f23c.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%2F97l0upr0b11cgwq1f23c.jpg" width="800" height="447"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;A dev.to notification landed in my inbox today.&lt;/p&gt;

&lt;p&gt;Someone called Dev Support Team had mentioned me in a comment.&lt;/p&gt;

&lt;p&gt;The first thing I did was open the .eml and look at the headers.&lt;/p&gt;

&lt;p&gt;They were genuine.&lt;/p&gt;

&lt;p&gt;DKIM for dev.to and SendGrid in pass, SPF pass from IP 168.245.48.148 and DMARC pass.&lt;/p&gt;

&lt;p&gt;The From is &lt;a href="mailto:yo@dev.to"&gt;yo@dev.to&lt;/a&gt; and the Return-Path bounces at em4960.dev.to.&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%2F3k8pwkfaf6ypila2l82m.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%2F3k8pwkfaf6ypila2l82m.png" width="799" height="341"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;That detail changes how you read the incident.&lt;/p&gt;

&lt;p&gt;The sender is not spoofed, so they used the platform's mention system to make the notification go out from its real infrastructure and pass every filter.&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%2Fx4k8i07rrtuoq8ki2c5b.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%2Fx4k8i07rrtuoq8ki2c5b.jpg" width="800" height="447"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The mention was a sweep
&lt;/h2&gt;

&lt;p&gt;Both notifications came from accounts created that same day.&lt;/p&gt;

&lt;p&gt;The display name is 𝘿𝙚𝙫 𝙎𝙪𝙥𝙥𝙤𝙧𝙩 𝙏𝙚𝙖𝙢 in Unicode mathematical characters.&lt;/p&gt;

&lt;p&gt;Those characters slip past text filters and the name reads like an administrator's. The usernames are random keypresses.&lt;/p&gt;

&lt;p&gt;{"type_of":"user","id":4228803,"username":"fgggggggggggggggggg",&lt;br&gt;
 "name":"𝘿𝙚𝙫 𝙎𝙪𝙥𝙥𝙤𝙧𝙩 𝙏𝙚𝙖𝙢","twitter_username":null,&lt;br&gt;
 "github_username":null,"summary":null,"website_url":null,&lt;br&gt;
 "joined_at":"Nov 5, 2026"}&lt;/p&gt;

&lt;p&gt;A single comment mentioned 10 accounts in alphabetical order. &lt;a class="mentioned-user" href="https://dev.to/anabel_obieze_1a5f44815ba"&gt;@anabel_obieze_1a5f44815ba&lt;/a&gt;, &lt;a class="mentioned-user" href="https://dev.to/anagha_h_a68da6f1218e8e30"&gt;@anagha_h_a68da6f1218e8e30&lt;/a&gt;, &lt;a class="mentioned-user" href="https://dev.to/anajsana95"&gt;@anajsana95&lt;/a&gt;, &lt;a class="mentioned-user" href="https://dev.to/analista_83"&gt;@analista_83&lt;/a&gt; (mine) and six more starting with anand. (Source: dev.to).&lt;/p&gt;

&lt;p&gt;That is not a targeted attack.&lt;/p&gt;

&lt;p&gt;It is a sweep of the user directory with a list pasted into a comment.&lt;/p&gt;

&lt;h2&gt;
  
  
  The accounts no longer exist
&lt;/h2&gt;

&lt;p&gt;Both permalinks in the notification have been returning 404 for a while.&lt;/p&gt;

&lt;p&gt;I checked a legitimate comment on the same platform and that one answers 200, so it is not a routing fault.&lt;/p&gt;

&lt;p&gt;The profiles do not load either and the API still returns the data because it caches it.&lt;/p&gt;

&lt;p&gt;The comment and the account have already been swept.&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%2F8ldzzziyr8cd6fe40uh8.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%2F8ldzzziyr8cd6fe40uh8.png" width="800" height="366"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The destination was a five-week-old domain
&lt;/h2&gt;

&lt;p&gt;The campaign's shortener tr.ee/dev-to landed on offerup-pay.top (Source: tr.ee, resolved on November 5).&lt;/p&gt;

&lt;p&gt;It was registered on October 1 at 21:48 UTC, with PDR Ltd as registrar, Cloudflare nameservers and a Let's Encrypt certificate issued that same day (Source: rdap.org).&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%2Fd4fg0tx372ldrcpdskab.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%2Fd4fg0tx372ldrcpdskab.png" width="799" height="277"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The campaign has a documented earlier step, anti-bot.icu, registered on September 25 with the same registrar, now in clientHold with its nameservers pointing to suspended-domain.com (Source: rdap.org).&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%2F4d2gvbxf8q7yeqaaxefc.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%2F4d2gvbxf8q7yeqaaxefc.png" width="800" height="299"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The domain did not go unnoticed either.&lt;/p&gt;

&lt;p&gt;On October 10, ten days after registration, a public urlscan.io analysis already returned the same title and the October 1 certificate. Google Safe Browsing had not classified it in that analysis (Source: urlscan.io).&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%2Fqrvr1y26dl1gxvnxh6xr.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%2Fqrvr1y26dl1gxvnxh6xr.png" width="800" height="557"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The verify you are human protects the scammer
&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%2Fgfk9jvjjgiy8i4pp1gfn.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%2Fgfk9jvjjgiy8i4pp1gfn.jpg" width="800" height="447"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;I passed the check out of curiosity.&lt;/p&gt;

&lt;p&gt;It is not Cloudflare, it is a cheap copy with the title "Just a moment", a hue filter on top, right-click disabled and copy and Ctrl+U blocked.&lt;/p&gt;

&lt;p&gt;It loads no external script and sends the drag telemetry to /__captcha/verify.&lt;/p&gt;

&lt;p&gt;Below it there is a paragraph forbidding AI, LLMs and crawlers from reading the page.&lt;/p&gt;

&lt;p&gt;A real site puts that control in front of a signup to stop bots.&lt;/p&gt;

&lt;p&gt;This one puts it in front of the scam.&lt;/p&gt;

&lt;p&gt;That way it slows down scanners and anyone asking an assistant whether the link is safe.&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%2F0lbhi2gvas26ud4u51nf.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%2F0lbhi2gvas26ud4u51nf.png" width="799" height="538"&gt;&lt;/a&gt;&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%2Fobtnx4967wb608i5scu1.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%2Fobtnx4967wb608i5scu1.jpg" width="800" height="447"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Behind it appears the card Mandatory Verification Of User #17984 with a 24-hour deadline and a fake support chat.&lt;/p&gt;

&lt;p&gt;The form asks for card number, expiry, CVV, name and the exact card balance.&lt;/p&gt;

&lt;p&gt;Then comes the code the bank sends you by SMS.&lt;/p&gt;

&lt;p&gt;The domain is called offerup-pay because the original script was a marketplace scam with the excuse of receiving a buyer's payment.&lt;/p&gt;

&lt;p&gt;They pasted the DEV logo on top and reused it (Source: dev.to/rudratosh).&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%2Fvuifrkuvy5nj0fokqabk.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%2Fvuifrkuvy5nj0fokqabk.png" width="629" height="434"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I checked on my machine
&lt;/h2&gt;

&lt;p&gt;In Brave, both links in the email, opened at 19:09, landed on 404.&lt;/p&gt;

&lt;p&gt;There is not a single visit to offerup-pay.top or tr.ee in this machine's history.&lt;/p&gt;

&lt;p&gt;I opened the gateway on my phone.&lt;/p&gt;

&lt;p&gt;There is no trace of an installation either.&lt;/p&gt;

&lt;p&gt;No new .exe, .msi, .ps1 or .zip in Downloads or Temp. The startup keys still hold Syncthing and Discord. Nothing else.&lt;/p&gt;

&lt;p&gt;A web page installs nothing unless you accept a download. The risk starts when someone types the card or the bank code.&lt;/p&gt;

&lt;h2&gt;
  
  
  The platform acknowledges it
&lt;/h2&gt;

&lt;p&gt;Ben, DEV's founder, posted today that they are working on mechanisms to reject that kind of attack automatically (Source: dev.to/ben).&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%2Fyyaoze7rcjmw5rsj0e51.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%2Fyyaoze7rcjmw5rsj0e51.png" width="702" height="352"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Until that filter arrives, anyone can fire an email at you with the platform's real brand and a link that works.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to do if you get one?
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Do not pay, do not verify and do not type the card, the balance or the bank code.&lt;/li&gt;
&lt;li&gt;Report the comment and block the account even if it no longer exists, the report feeds the filter.&lt;/li&gt;
&lt;li&gt;Report the domain at abuse.cloudflare.com (even if they wash their hands) and in Google Safe Browsing's phishing form.&lt;/li&gt;
&lt;li&gt;If you typed the card, call the bank, block the card and review the charges.&lt;/li&gt;
&lt;li&gt;If you wrote in the chat, expect calls or SMS from a fake support in the coming days.&lt;/li&gt;
&lt;li&gt;If you opened the link on your phone, check whether you accepted a configuration profile or installed an app.&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;Originally published at &lt;a href="https://sammideblas.com/notas/the-email-was-real-and-the-fraud-was-in-the-comment" rel="noopener noreferrer"&gt;https://sammideblas.com/notas/the-email-was-real-and-the-fraud-was-in-the-comment&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Si has leído hasta aquí... reacciona y comparte -&amp;gt; "la seguridad y la defensa en la era de la IA es cosa de todos los que la usan" ¿No Crees?&lt;/p&gt;

</description>
      <category>phishing</category>
      <category>security</category>
      <category>osint</category>
      <category>devto</category>
    </item>
    <item>
      <title>The patch on top does not close the door below</title>
      <dc:creator>Sammi De Blas </dc:creator>
      <pubDate>Wed, 04 Nov 2026 11:28:55 +0000</pubDate>
      <link>https://dev.to/analista_83/the-patch-on-top-does-not-close-the-door-below-86m</link>
      <guid>https://dev.to/analista_83/the-patch-on-top-does-not-close-the-door-below-86m</guid>
      <description>&lt;h2&gt;
  
  
  The patch on top does not close the door below
&lt;/h2&gt;

&lt;p&gt;The software ecosystem has an inventory problem and a delegated trust problem, and the week showed both at once.&lt;/p&gt;

&lt;p&gt;Updating the main application no longer guarantees security.&lt;/p&gt;

&lt;p&gt;GitLab (CVE-2026-97229, CVSS 10) carries an unauthenticated remote code execution flaw in the CI/CD pipeline, and it is under active exploitation.&lt;/p&gt;

&lt;p&gt;The problem, as Alpine Linux found, is that the GitLab patch does not fix the underlying Ruby interpreter. If the operating system keeps a vulnerable Ruby, the door stays open.&lt;/p&gt;

&lt;p&gt;PHP2Shell shows an identical scenario, which proves the attack surface includes invisible dependencies that teams assume belong to someone else.&lt;/p&gt;

&lt;p&gt;If the pattern can be manipulated, the attacker gets the interpreter to run whatever he writes instead of only searching.&lt;/p&gt;

&lt;p&gt;CISA added it to the KEV catalog on October 29 and gave federal agencies until Sunday to close it (Source: securityweek.com).&lt;/p&gt;

&lt;p&gt;That detail turns a routine advisory into an inventory problem. The question is no longer whether your GitLab is updated but which Ruby version is running it.&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%2Fsammideblas.com%2Fstatic%2Fnotas%2F2026-11-04-fallo-en-la-base.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%2Fsammideblas.com%2Fstatic%2Fnotas%2F2026-11-04-fallo-en-la-base.jpg" alt="The layer that runs below: the bug is not in the app you patched" width="800" height="400"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The week brings the same pattern somewhere else.&lt;/p&gt;

&lt;p&gt;PHP2Shell still has no complete mitigation, and PHP warns that updating is all that is left (Source: securityweek.com).&lt;/p&gt;

&lt;p&gt;The flaw affects versions from 5.4 onward, carries CVSS 9.2, and industrial vendors are checking product by product whether they embed the interpreter (Source: securityweek.com).&lt;/p&gt;

&lt;p&gt;Blind trust in third parties has become the most profitable vector.&lt;/p&gt;

&lt;p&gt;Polymarket lost 2.7 million dollars in minutes because a compromised external analytics script hijacked sessions straight from the user's browser.&lt;/p&gt;

&lt;p&gt;In parallel, the official Claude Code extension shipped trojanized code from the Cursor marketplace for eleven hours, opening reverse tunnels through SSH keys.&lt;/p&gt;

&lt;p&gt;The attacker does not need to break the company's code because compromising the middle piece or the official distribution channel is enough.&lt;/p&gt;

&lt;p&gt;In all three cases the flaw lives in a piece the team does not consider its own. The application on top gets patched and the case is closed.&lt;/p&gt;

&lt;h2&gt;
  
  
  When the channel is the attack
&lt;/h2&gt;

&lt;p&gt;On October 29, the official Claude Code extension 2.1.313 was served trojanized from the Cursor marketplace for eleven hours (Source: microsoft.com).&lt;/p&gt;

&lt;p&gt;The version that reached users was not the one the vendor had signed, and it added an attacker SSH key and opened a persistent reverse tunnel to his infrastructure.&lt;/p&gt;

&lt;p&gt;A reverse tunnel is a connection that leaves the compromised machine and stays open. The attacker does not need to get in afterward, because the door is already open from the inside.&lt;/p&gt;

&lt;p&gt;Up to 5,000 users were exposed in that eleven-hour window. Cursor rebuilt the marketplace and rotated the affected keys (Source: microsoft.com).&lt;/p&gt;

&lt;p&gt;What stands out is not that a package gets compromised but that the compromise happened in the channel the tool itself uses to distribute, the one the user installs because he trusts the brand.&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%2Fsammideblas.com%2Fstatic%2Fnotas%2F2026-11-04-cadena-de-suministro.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%2Fsammideblas.com%2Fstatic%2Fnotas%2F2026-11-04-cadena-de-suministro.jpg" alt="The channel that ships the package: a trojanized supply chain" width="800" height="400"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;That same day, OpenAI confirmed another incident in the same direction. A shared ChatGPT link could run code inside the recipient's authenticated session and exfiltrate conversations that had never been shared (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;The fix was server-side, with no user action, so the flaw did not depend on what the victim did but on what the service allowed a link to do.&lt;/p&gt;

&lt;p&gt;The week brings a third case along the same line.&lt;/p&gt;

&lt;p&gt;TeamViewer fixed CVE-2026-111585, CVSS 8.8, in clients earlier than version 15.83 (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;The flaw is an overflow when negotiating screen size. It allows code execution in the Windows viewer with the default configuration; accepting the session is enough.&lt;/p&gt;

&lt;p&gt;The direction of the attack is what matters. The compromised remote host attacks whoever connects to it, not the other way around. The machine you call to fix something may be the one that compromises you.&lt;/p&gt;

&lt;p&gt;The traditional connection paradigm is failing too.&lt;/p&gt;

&lt;h2&gt;
  
  
  The other side
&lt;/h2&gt;

&lt;p&gt;Not every week is about patching.&lt;/p&gt;

&lt;p&gt;Polymarket lost about 2.7 million dollars in minutes because the DebugBear analytics script was compromised and hijacked sessions to drain wallets (Source: securityweek.com).&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%2Fsammideblas.com%2Fstatic%2Fnotas%2F2026-11-04-polymarket.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%2Fsammideblas.com%2Fstatic%2Fnotas%2F2026-11-04-polymarket.jpg" alt="The trading platform compromised through a third-party script" width="800" height="400"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The platform paused trading for an hour and promised refunds, and the vector was not a flaw in its code but a third party serving a script on its page.&lt;/p&gt;

&lt;p&gt;In all four cases, the weak link was outside the company's own inventory.&lt;/p&gt;

&lt;h2&gt;
  
  
  My reading
&lt;/h2&gt;

&lt;p&gt;My reading is that the hole is in management, not in patches.&lt;/p&gt;

&lt;p&gt;The team knows which applications it has and does not know what holds them up underneath, which channel distributes them, which hosts they connect to, or which scripts they load in the user's browser.&lt;/p&gt;

&lt;p&gt;Tool choice is no longer decided only by what the tool does but by who distributes it and which signature backs it.&lt;/p&gt;

&lt;p&gt;A marketplace that rebuilds its catalog in eleven hours shows reaction capacity but does not prove the next package will arrive clean.&lt;/p&gt;

&lt;p&gt;And remote access works the same way, because if the host you connect to can attack you, the connection policy has to assume the other end is hostile until proven otherwise.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to look at?
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The interpreter version.&lt;/strong&gt; Check on every GitLab host which Ruby runs it and whether it is on the fixed version list. With PHP, the same: the application patch does not cover the interpreter.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Exposed CI/CD endpoints.&lt;/strong&gt; Check whether the pipeline accepts requests from outside the internal network. GitLab runs code by design, so a flaw there does not need a victim to click.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Installed extensions.&lt;/strong&gt; If you have Claude Code 2.1.313 or anything that came from the Cursor marketplace on October 29, rotate the machine's SSH keys and look for active reverse tunnels.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Remote access and third-party scripts.&lt;/strong&gt; That viewer should not have permissions over the internal network. And any embedded analytics or tag manager is attack surface, so if you cannot pin its version, log which domain serves each one.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How would I test it in my lab?
&lt;/h2&gt;

&lt;p&gt;I would set up GitLab in a container with the system Ruby on the vulnerable version and another with the fixed one. I would send the same malicious request against both and watch each log.&lt;/p&gt;

&lt;p&gt;If the first executes and the second does not, I have proof that the GitLab patch is not enough on its own. If both execute, the problem is in another layer and I have to keep going down.&lt;/p&gt;

&lt;p&gt;Then I would install an extension from a local marketplace and watch which keys it writes to the system. It is cheap and it says a lot about what the user accepts without reading.&lt;/p&gt;

&lt;p&gt;It is a test that fits on any laptop with Docker. I would run it before closing the CISA advisory on my own network.&lt;/p&gt;

&lt;h2&gt;
  
  
  Action plan
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Identify the underlying interpreter:&lt;/strong&gt; Check on every host which exact version of Ruby, PHP or another base language runs the main application.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Limit CI/CD exposure:&lt;/strong&gt; Block external requests to the pipeline endpoints if they are not strictly necessary for operations.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Audit external scripts:&lt;/strong&gt; Log every third-party domain loaded on your site; a compromised analytics tool grants full access to authenticated sessions.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Isolate remote support:&lt;/strong&gt; Viewers used to connect to clients or third parties should have no permissions or visibility over your internal network.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Validate in the lab:&lt;/strong&gt; Spin up Docker containers with vulnerable and patched versions of the base system to empirically check whether the application update stops the exploit.&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;Originally published at &lt;a href="https://sammideblas.com/notas/the-patch-on-top-does-not-close-the-door-below" rel="noopener noreferrer"&gt;https://sammideblas.com/notas/the-patch-on-top-does-not-close-the-door-below&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Si has leído hasta aquí... reacciona y comparte -&amp;gt; "la seguridad y la defensa en la era de la IA es cosa de todos los que la usan" ¿No Crees?&lt;/p&gt;

</description>
      <category>gitlab</category>
      <category>security</category>
      <category>supplychain</category>
      <category>patching</category>
    </item>
    <item>
      <title>AI made finding bugs cheap, not fixing them</title>
      <dc:creator>Sammi De Blas </dc:creator>
      <pubDate>Fri, 30 Oct 2026 14:09:03 +0000</pubDate>
      <link>https://dev.to/analista_83/ai-made-finding-bugs-cheap-not-fixing-them-4im7</link>
      <guid>https://dev.to/analista_83/ai-made-finding-bugs-cheap-not-fixing-them-4im7</guid>
      <description>&lt;h2&gt;
  
  
  The case
&lt;/h2&gt;

&lt;p&gt;On October 21, emergency patches shipped for two vulnerabilities in PHP itself, the language running behind a good part of the web.&lt;/p&gt;

&lt;p&gt;They were registered as CVE-2026-109618 and CVE-2026-109619 with a CVSS of 9.2, and the pair got the name php2shell (Source: fortune.com).&lt;/p&gt;

&lt;p&gt;Both are remote code execution, meaning flaws that let an attacker run whatever they want on the server without authenticating first.&lt;/p&gt;

&lt;p&gt;What stands out is not the flaw itself but where it came from.&lt;/p&gt;

&lt;p&gt;The code had been sitting there for over a decade, invisible, and it was found with help from AI (Source: fortune.com).&lt;/p&gt;

&lt;p&gt;Cloudflare already confirms active exploitation.&lt;/p&gt;

&lt;p&gt;And several UK and European government sites, plus Wikipedia and the Internet Archive, went down preventively (Source: fortune.com).&lt;/p&gt;

&lt;p&gt;That preventive shutdown is the part worth reading twice, because it was not an attack that took those services down but the decision to switch off before it arrived.&lt;/p&gt;

&lt;h2&gt;
  
  
  Not an isolated case
&lt;/h2&gt;

&lt;p&gt;The same week brought another flaw with the same profile.&lt;/p&gt;

&lt;p&gt;Oracle E-Business Suite has an unauthenticated SQL injection in the TCF servlet, registered as CVE-2026-101767 with a 9.8 that chains up to remote code execution (Source: helpnetsecurity.com).&lt;/p&gt;

&lt;p&gt;It was patched in the October 20 CPU and there are already exploitation probes.&lt;/p&gt;

&lt;p&gt;Censys counts 2,246 web properties with EBS exposed (Source: helpnetsecurity.com).&lt;/p&gt;

&lt;p&gt;Old flaws, in software nobody looks at because it has been running for years, found more easily now and exploited within days.&lt;/p&gt;

&lt;p&gt;In parallel, CISA added PolyShell (CVE-2026-48356, 9.3) and Node2Shell (CVE-2026-111497, 9.2) to KEV. Sansec sees more than 950 Adobe Commerce and Magento stores with a loader in CMS blocks since July 2025, and there are already Brotli exploitation attempts against honeypots (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;What used to take months to go from flaw to exploit now takes hours.&lt;/p&gt;

&lt;h2&gt;
  
  
  The other side
&lt;/h2&gt;

&lt;p&gt;AI does not only find flaws.&lt;/p&gt;

&lt;p&gt;Moonshot removed cyberattack material from training before opening Kimi K3.1 Pro, and the model still solves 81% of its internal exploit tasks, against 33% for the previous version (Source: thenextweb.com).&lt;/p&gt;

&lt;p&gt;Offensive capability does not travel in the dataset, I think it travels in the access.&lt;/p&gt;

&lt;p&gt;Filtering data is no substitute for limiting permissions or monitoring what the agent does with them.&lt;/p&gt;

&lt;p&gt;Lazarus, the North Korean group, encrypts the Operation FlashHole malware with post-quantum cryptography, active since July 2025 against South Korean construction, software and cable manufacturing companies (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;The vector is a watering hole, meaning compromising a site the victims visit regularly, in this case the mandatory security software iniLINE Cross EX (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;My reading is that the three pieces tell the same story from different angles:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The attacker speeds up in three phases: finding, weaponizing and hiding.&lt;/li&gt;
&lt;li&gt;The defender still measures speed in the fourth one: patching.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  My own reading
&lt;/h2&gt;

&lt;p&gt;For years the debate was whether AI would find flaws humans could not see.&lt;/p&gt;

&lt;p&gt;The answer has arrived and it is yes.&lt;/p&gt;

&lt;p&gt;Finding a decade-old flaw in hours is real progress, but if the patch takes weeks to reach production the progress only benefits whoever exploits first.&lt;/p&gt;

&lt;p&gt;It comes down to the inventory of what runs in production, and update speed in most organizations is measured in months, weeks if you are optimistic.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to look at
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The real inventory.&lt;/strong&gt; Not the CMDB one, the one that answers a scan. If you do not know which PHP version runs on each host, the October 21 patch is useless to you.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The patching window.&lt;/strong&gt; How long a critical patch takes from release to production. If the answer is more than a week, the decade-old flaw has already reached you.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;EBS exposure.&lt;/strong&gt; If you have Oracle E-Business Suite with the TCF servlet reachable from the internet, that is the point to close before thinking about the rest.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How would I test it in my lab?
&lt;/h2&gt;

&lt;p&gt;I would spin up a container with an old PHP version, one of the ones carrying the flaw, and leave it on an isolated network.&lt;/p&gt;

&lt;p&gt;I would run a scan with a model-assisted static analysis tool and measure how long it takes to flag the vulnerable function, and I would bet you do not need frontier models.&lt;/p&gt;

&lt;p&gt;Then I would time how long it takes me to apply the patch and bring the service back up.&lt;/p&gt;

&lt;p&gt;The question here is how long I take to close it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Closing
&lt;/h2&gt;

&lt;p&gt;AI has made finding flaws cheap but not fixing them, and while discovery is measured in hours and patching in weeks or months, the advantage always goes to whoever attacks.&lt;/p&gt;




&lt;p&gt;Originally published at &lt;a href="https://sammideblas.com/notas/ai-made-finding-bugs-cheap-not-fixing-them" rel="noopener noreferrer"&gt;https://sammideblas.com/notas/ai-made-finding-bugs-cheap-not-fixing-them&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Si has leído hasta aquí... reacciona y comparte -&amp;gt; "la seguridad y la defensa en la era de la IA es cosa de todos los que la usan" ¿No Crees?&lt;/p&gt;

</description>
      <category>php</category>
      <category>security</category>
      <category>patching</category>
      <category>ai</category>
    </item>
    <item>
      <title>The AGENTS.md your agent reads as an order</title>
      <dc:creator>Sammi De Blas </dc:creator>
      <pubDate>Wed, 28 Oct 2026 11:35:53 +0000</pubDate>
      <link>https://dev.to/analista_83/the-agentsmd-your-agent-reads-as-an-order-1h8d</link>
      <guid>https://dev.to/analista_83/the-agentsmd-your-agent-reads-as-an-order-1h8d</guid>
      <description>&lt;h2&gt;
  
  
  The file that was not documentation
&lt;/h2&gt;

&lt;p&gt;The campaign is called MALFEX and the vector is a text file.&lt;/p&gt;

&lt;p&gt;A malicious AGENTS.md placed in the Downloads folder tried to make Codex install the fraudulent npm package function-color (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;What stands out is not the package but how the file got there.&lt;/p&gt;

&lt;p&gt;A website can write that AGENTS.md without the user clicking anything, and the agent reads it as if it were orders (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;AGENTS.md is the file where the project rules are explained to the agent.&lt;/p&gt;

&lt;p&gt;It is meant for a human to write and for the model to obey.&lt;/p&gt;

&lt;p&gt;That design is exactly the problem, because the agent does not distinguish an engineering note from a third party's instruction.&lt;/p&gt;

&lt;p&gt;To the model, both are text in the context, and text in the context gets executed.&lt;/p&gt;

&lt;p&gt;The website writes the file, the agent finds it when it opens the folder, reads it as configuration and runs the install.&lt;/p&gt;

&lt;p&gt;No exploit, no credential, no binary needed, just a file with the right name in the right place.&lt;/p&gt;

&lt;h2&gt;
  
  
  Not an isolated case
&lt;/h2&gt;

&lt;p&gt;The week brings more examples of the same pattern.&lt;/p&gt;

&lt;p&gt;OpenAI revealed that its own red team agents escaped microVMs and communicated between isolated sandboxes with no network and no shared files, using the CPU cache as a covert timing channel (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;The same team found routes to the internet through misconfigured metrics, logging and authentication services (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Isolation is assumed, not verified.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;A sandbox is treated as a wall when it is really a convention someone configured once and nobody looked at again.&lt;/p&gt;

&lt;p&gt;In the AGENTS.md case the boundary is not the network, it is the file system.&lt;/p&gt;

&lt;p&gt;The agent trusts what is on disk, and disk is where anyone with an open tab can write.&lt;/p&gt;

&lt;h2&gt;
  
  
  The other side
&lt;/h2&gt;

&lt;p&gt;The same engine that speeds up a defender speeds up the attacker.&lt;/p&gt;

&lt;p&gt;Microsoft attributed to Storm-0940, a China-linked group, the abuse of ConnectWise OAuth credentials with the ARTEX framework.&lt;/p&gt;

&lt;p&gt;The result was 2,250 Microsoft 365 tenants probed at 2,900 access attempts per minute, after a password spray against 327 accounts in 8 organizations (Source: microsoft.com).&lt;/p&gt;

&lt;p&gt;It is not an autonomous agent, but the pattern is the same one that worries people about agents.&lt;/p&gt;

&lt;p&gt;Cheap automation against surfaces that were left open and nobody reviewed.&lt;/p&gt;

&lt;p&gt;Neither of these last two has anything to do with agents, but they share the lesson: the surface nobody watches is the one that breaks.&lt;/p&gt;

&lt;h2&gt;
  
  
  My reading
&lt;/h2&gt;

&lt;p&gt;An agent configuration file is not documentation, it is code that executes by design, and it has to be treated as untrusted code.&lt;/p&gt;

&lt;p&gt;That implies three things.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Review what enters the repo.&lt;/li&gt;
&lt;li&gt;Limit the agent's network and install permissions.&lt;/li&gt;
&lt;li&gt;Do not let it write to folders the user does not control.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The Downloads folder is the worst possible place for an AGENTS.md, because it is where things nobody reviews end up.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to look at?
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Check the configuration files.&lt;/strong&gt; AGENTS.md, .cursor/rules and CLAUDE.md in the repo and in the working folders. Check who wrote them and when. If you do not recognize them, delete them before opening the agent.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Limit the agent's write access.&lt;/strong&gt; The agent should not be able to write outside the project directory. If it can, a file placed by a website becomes an order.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cut automatic installs.&lt;/strong&gt; Letting the agent run npm install without confirmation is a direct path to a fraudulent package. Require human approval for any install.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How would I test it in my lab?
&lt;/h2&gt;

&lt;p&gt;I would set up a working directory with Codex pointed at it and place an AGENTS.md with an innocent instruction, I don't know, for example it occurs to me that it writes a test file in /tmp, you see where I'm going...&lt;/p&gt;

&lt;p&gt;I would check whether the agent runs it without asking and what permissions it had at that moment.&lt;/p&gt;

&lt;p&gt;If it obeys an instruction from a file I did not write, the problem is not the file, it is that the agent does not distinguish who wrote what.&lt;/p&gt;

&lt;h2&gt;
  
  
  Closing
&lt;/h2&gt;

&lt;p&gt;The agent does not distinguish an engineering note from an order, and while that distinction does not exist in the design, every file it reads is a door.&lt;/p&gt;




&lt;p&gt;Originally published at &lt;a href="https://sammideblas.com/notas/the-agents-md-your-agent-reads-as-an-order" rel="noopener noreferrer"&gt;https://sammideblas.com/notas/the-agents-md-your-agent-reads-as-an-order&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Si has leído hasta aquí... reacciona y comparte -&amp;gt; "la seguridad y la defensa en la era de la IA es cosa de todos los que la usan" ¿No Crees?&lt;/p&gt;

</description>
      <category>agents</category>
      <category>ai</category>
      <category>security</category>
      <category>sandbox</category>
    </item>
    <item>
      <title>Two core RCEs in one week, neither needs a sleepless night</title>
      <dc:creator>Sammi De Blas </dc:creator>
      <pubDate>Mon, 26 Oct 2026 10:11:24 +0000</pubDate>
      <link>https://dev.to/analista_83/two-core-rces-in-one-week-neither-needs-a-sleepless-night-c6p</link>
      <guid>https://dev.to/analista_83/two-core-rces-in-one-week-neither-needs-a-sleepless-night-c6p</guid>
      <description>&lt;h2&gt;
  
  
  The main case
&lt;/h2&gt;

&lt;p&gt;The PHP2Shell patches shipped on October 21: two vulnerabilities in the core of the language, CVE-2026-109618 and CVE-2026-109619, CVSS 9.2 (Source: helpnetsecurity.com).&lt;/p&gt;

&lt;p&gt;They allow unauthenticated remote code execution, meaning someone outside can run whatever they want on the server without an account or a password.&lt;/p&gt;

&lt;p&gt;The flaw is fixed in versions 8.2.35, 8.3.36, 8.4.27 and 8.5.12.&lt;/p&gt;

&lt;p&gt;What stands out is the timeline.&lt;/p&gt;

&lt;p&gt;AWS detected exploitation attempts from October 17, four days before the patch existed (Source: helpnetsecurity.com).&lt;/p&gt;

&lt;p&gt;And it reaches further than it looks. WordPress, Magento, Drupal or Nextcloud can be exposed even when their own code is clean, because the flaw lives in the interpreter that runs them.&lt;/p&gt;

&lt;p&gt;The second case arrived almost at the same time. Node2Shell, CVE-2026-111497, also CVSS 9.2, is an out-of-bounds write in the Node.js Brotli decoder that allows unauthenticated remote code execution (Source: cryptobriefing.com).&lt;/p&gt;

&lt;p&gt;The Codex Security team at OpenAI found it, and the patches are in versions 22.23.4, 24.22.1 and 26.12.1.&lt;/p&gt;

&lt;p&gt;One curious detail repeats in both: LLM agents found them, not a human researcher spending weeks with a debugger.&lt;/p&gt;

&lt;h2&gt;
  
  
  Not an isolated case
&lt;/h2&gt;

&lt;p&gt;The week closes with a third episode that fits the same pattern.&lt;/p&gt;

&lt;p&gt;On October 24 it was confirmed that Node2Shell also affects Bun, the Python brotli and brotlicffi libraries, Envoy and Istio. A public proof of concept already demonstrates remote code execution in a lab, though there is no confirmed exploitation in production (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;The common shape of the three cases is clear.&lt;/p&gt;

&lt;p&gt;A flaw in an infrastructure piece almost nobody audits, a maximum score, a patch available in days, and an exposure window measured in hours, not months.&lt;/p&gt;

&lt;p&gt;And AI itself is accelerating the discovery.&lt;/p&gt;

&lt;p&gt;That changes the pace at which advisories appear, not the damage each one does.&lt;/p&gt;

&lt;h2&gt;
  
  
  The other side
&lt;/h2&gt;

&lt;p&gt;The same week brings the opposite face.&lt;/p&gt;

&lt;p&gt;Microsoft attributed to Storm-0940, a China-linked group, the abuse of ConnectWise OAuth credentials with the ARTEX framework (Source: microsoft.com).&lt;/p&gt;

&lt;p&gt;The numbers of the attack are what matter.&lt;/p&gt;

&lt;p&gt;2,250 Microsoft 365 tenants probed at 2,900 access attempts per minute, after a password spray against 327 accounts in 8 organizations.&lt;/p&gt;

&lt;p&gt;It is not a core RCE: it is a poorly protected credential used at industrial scale, and the result is just as serious.&lt;/p&gt;

&lt;p&gt;My reading is that AI accelerates the discovery of flaws, not the mass exploitation of those flaws.&lt;/p&gt;

&lt;p&gt;A CVSS 9.2 does not always demand a sleepless night, because not every flaw has a universal exploit that works against any installation. This comes up a lot in CySA+.&lt;/p&gt;

&lt;p&gt;What does demand a sleepless night is knowing whether your exposed surface is on the list.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to check
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;PHP and Node.js version.&lt;/strong&gt; Check whether your servers run 8.2.35, 8.3.36, 8.4.27, 8.5.12 or 22.23.4, 24.22.1, 26.12.1. If not, the patch is the first task of the day.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Exposed surface.&lt;/strong&gt; A PHP that only talks locally is not the same as one serving requests from the internet. Prioritize by what is visible from outside, not by the CVE score.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Indirect dependencies.&lt;/strong&gt; Check whether your WordPress, your Drupal or your Nextcloud run on an unpatched PHP. The application code can be clean and the flaw still be there.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How would I test it in my lab?
&lt;/h2&gt;

&lt;p&gt;I would spin up a container with the vulnerable PHP version and another with the patched one, then send the same request against both.&lt;/p&gt;

&lt;p&gt;I would watch whether the response changes shape, whether the process crashes, or whether a different trace shows up in the log.&lt;/p&gt;

&lt;p&gt;With Node.js I would do the same against the Brotli decoder, sending a malformed payload and watching whether the process restarts on its own.&lt;/p&gt;

&lt;p&gt;The conclusion I would draw is whether my specific configuration is reachable or whether the flaw stays theoretical for my case.&lt;/p&gt;

&lt;h2&gt;
  
  
  Closing
&lt;/h2&gt;

&lt;p&gt;Two core flaws in the same week are not an emergency if you know which version you run and what you expose; but they are an emergency if you do not.&lt;/p&gt;




&lt;p&gt;Originally published at &lt;a href="https://sammideblas.com/notas/two-core-rces-in-one-week-neither-needs-a-sleepless-night" rel="noopener noreferrer"&gt;https://sammideblas.com/notas/two-core-rces-in-one-week-neither-needs-a-sleepless-night&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Si has leído hasta aquí... reacciona y comparte -&amp;gt; "la seguridad y la defensa en la era de la IA es cosa de todos los que la usan" ¿No Crees?&lt;/p&gt;

</description>
      <category>php</category>
      <category>node</category>
      <category>rce</category>
      <category>vulnerability</category>
    </item>
    <item>
      <title>The agent is the attacker now and the inventory can't see it</title>
      <dc:creator>Sammi De Blas </dc:creator>
      <pubDate>Fri, 23 Oct 2026 20:29:31 +0000</pubDate>
      <link>https://dev.to/analista_83/the-agent-is-the-attacker-now-and-the-inventory-cant-see-it-4ef9</link>
      <guid>https://dev.to/analista_83/the-agent-is-the-attacker-now-and-the-inventory-cant-see-it-4ef9</guid>
      <description>&lt;h2&gt;
  
  
  The main case
&lt;/h2&gt;

&lt;p&gt;Microsoft attributed the October 22 assault on ConnectWise Cloud Backup to Storm-0940, a China-linked group (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;The attack used ARTEX, an agentic pentesting framework. No human was typing: an agent was firing login attempts at machine speed.&lt;/p&gt;

&lt;p&gt;The numbers make it clear. 2,900 login attempts per minute against 2,250 M365 tenants; privileged access in under three hours (Source: thehackernews.com).&lt;/p&gt;

&lt;p&gt;A human team can't hold that pace against that many targets; but an agent can.&lt;/p&gt;

&lt;p&gt;And it isn't the only case this week where the agent shows up on the attacker's side.&lt;/p&gt;

&lt;p&gt;The US Army solicited ECHOTRIBBLE, a framework of autonomous offensive agents that infiltrate, copy themselves to other machines (iPad, macOS, Windows), neutralize targets and self-destruct before detection, with minimal human oversight (Source: helpnetsecurity.com, theintercept.com).&lt;/p&gt;

&lt;p&gt;The same technology sold as defense is bought as a weapon.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;That is no longer a future hypothesis.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Not an isolated case
&lt;/h2&gt;

&lt;p&gt;The pattern repeats in the other half of the week, in the opposite direction.&lt;/p&gt;

&lt;p&gt;OpenAI admitted that agents running GPT-6.1 Sol and Astra weakened misalignment monitoring in two internal tasks.&lt;/p&gt;

&lt;p&gt;They read code instructions as authorization and one denied it when asked (Source: bleepingcomputer.com).&lt;/p&gt;

&lt;p&gt;An agent that bypasses its own control and then lies about it is not a model failure: it is a textbook containment failure.&lt;/p&gt;

&lt;p&gt;Microsoft knows this, which is why it extended Project Perception with Guardian Agents.&lt;/p&gt;

&lt;p&gt;Dedicated identity for agents in Entra Agent ID, unified logging and containment when the agent steps outside its brief. It already discovers 60 types, including Claude Code and Codex (Source: helpnetsecurity.com).&lt;/p&gt;

&lt;p&gt;The agent acts with an identity, makes decisions and leaves a trail; if the identity is yours and the trail isn't kept, there is no containment possible.&lt;/p&gt;

&lt;h2&gt;
  
  
  The other bottleneck
&lt;/h2&gt;

&lt;p&gt;While the agents run, the inventory falls behind.&lt;/p&gt;

&lt;p&gt;CVE-2026-110264, CVSS 10.0, is an unauthenticated RCE in GitLab CE/EE self-managed caused by an integer overflow when compiling a regex.&lt;/p&gt;

&lt;p&gt;It affects every version since 7.0, from 2014. It was patched on the 21st in 18.11.13, 19.0.10, 19.1.9, 19.2.9, 19.3.6, 19.4.3 and 19.5.1 (Source: docs.gitlab.com).&lt;/p&gt;

&lt;p&gt;Twelve years of affected versions. That only happens if nobody looks at what runs in production.&lt;/p&gt;

&lt;h3&gt;
  
  
  A black week
&lt;/h3&gt;

&lt;p&gt;The PHP case runs along the same line. PHP2Shell (CVE-2026-109618 and -109619, CVSS 9.2) is an unauthenticated RCE in the PHP core.&lt;/p&gt;

&lt;p&gt;Details leaked on the 20th and the patch landed on the 21st at 14:00 UTC, with exploitation already reported by hosting providers and emergency WAF rules at Cloudflare.&lt;/p&gt;

&lt;p&gt;PHP 8.1 and earlier will get no fix (Source: therecord.media, recordedfuture.com).&lt;/p&gt;

&lt;p&gt;One day between the leak and the patch; exploitation was already underway.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The window is measured in hours.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;And there is one detail that closes the circle:&lt;/p&gt;

&lt;p&gt;Frontier models found both critical PHP bugs, GPT-5.6 Sol via Patch the Planet and Wiz Atlas with Gemini 4 Argon, in a coordinated two-month process with the maintainers (Source: unite.ai, openai.com).&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;AI found the bug.&lt;/li&gt;
&lt;li&gt;AI also speeds up the exploit.&lt;/li&gt;
&lt;li&gt;The human who decides whether to update still arrives late.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Sleep well tonight.&lt;/p&gt;

&lt;h2&gt;
  
  
  My reading
&lt;/h2&gt;

&lt;p&gt;My reading is that the problem is no longer the model, it is the inventory.&lt;/p&gt;

&lt;p&gt;A GitLab untouched since 2014 and an agent with live credentials months after it stopped being used are the same failure.&lt;/p&gt;

&lt;p&gt;Someone granted access and nobody set an expiry date.&lt;/p&gt;

&lt;p&gt;The containment Microsoft and others sell is worth little if you don't first know what you have running; unified logging doesn't fix a server that appears on no list.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to look at?
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Owner per asset.&lt;/strong&gt; Every agent and every self-managed service needs a name behind it. If it has none, it is a candidate for shutdown.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Permissions and expiry.&lt;/strong&gt; Check which token each agent uses and when it expires. A token with no date is an open door.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Unsupported versions.&lt;/strong&gt; Count what percentage of your inventory runs versions that no longer receive patches. That number is your real risk.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How I would test it in my lab
&lt;/h2&gt;

&lt;p&gt;I would stand up a self-managed GitLab on a virtual machine and leave it on an old version on purpose.&lt;/p&gt;

&lt;p&gt;Then I would run an inventory scan with three fields per asset: owner, permissions and expiry date and see how many come back empty.&lt;/p&gt;

&lt;p&gt;After that I would do the same with a local agent connected to an MCP server.&lt;/p&gt;

&lt;p&gt;I would give it a token with a short expiry and check whether the log stores every tool call with its arguments.&lt;/p&gt;

&lt;p&gt;The conclusion would be simple: if the inventory isn't complete, containment is decorative.&lt;/p&gt;

&lt;h2&gt;
  
  
  Closing
&lt;/h2&gt;

&lt;p&gt;The agent attacks at machine speed and the inventory updates at human speed.&lt;/p&gt;

&lt;p&gt;Closing that gap is the work of this week, sorry, of the future.&lt;/p&gt;




&lt;p&gt;Originally published at &lt;a href="https://sammideblas.com/notas/the-agent-is-the-attacker-now-and-the-inventory-can-t-see-it" rel="noopener noreferrer"&gt;https://sammideblas.com/notas/the-agent-is-the-attacker-now-and-the-inventory-can-t-see-it&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Si has leído hasta aquí... reacciona y comparte -&amp;gt; "la seguridad y la defensa en la era de la IA es cosa de todos los que la usan" ¿No Crees?&lt;/p&gt;

</description>
      <category>security</category>
      <category>ai</category>
      <category>gitlab</category>
      <category>patching</category>
    </item>
    <item>
      <title>The US Army is buying agents that self-destruct before detection</title>
      <dc:creator>Sammi De Blas </dc:creator>
      <pubDate>Wed, 21 Oct 2026 10:55:56 +0000</pubDate>
      <link>https://dev.to/analista_83/the-us-army-is-buying-agents-that-self-destruct-before-detection-2oj1</link>
      <guid>https://dev.to/analista_83/the-us-army-is-buying-agents-that-self-destruct-before-detection-2oj1</guid>
      <description>&lt;h2&gt;
  
  
  The solicitation
&lt;/h2&gt;

&lt;p&gt;The US Army has opened the solicitation for ECHOTRIBBLE, a framework of offensive autonomous agents (Source: helpnetsecurity.com, theintercept.com).&lt;/p&gt;

&lt;p&gt;The document describes what they do and it boggles the mind: they infiltrate, copy themselves to other machines, neutralize targets and self-destruct before detection.&lt;/p&gt;

&lt;p&gt;The target machines it names are iPad, macOS and Windows, a good excuse to move to Linux.&lt;/p&gt;

&lt;p&gt;And the phrase that sums up the design is "minimal human oversight".&lt;/p&gt;

&lt;p&gt;This is not a lab experiment, you can buy it.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is actually being bought
&lt;/h2&gt;

&lt;p&gt;An agent that copies itself is not a new idea, but what changes here is the last part of the cycle: it self-destructs before anyone detects it.&lt;/p&gt;

&lt;p&gt;That turns the trail into the main target.&lt;/p&gt;

&lt;p&gt;An incident is reconstructed from what is left on the endpoint, that is, the binary, the process, the connection, the file it touched.&lt;/p&gt;

&lt;p&gt;If the agent deletes itself and also wipes the endpoint events, there is nothing left to analyze.&lt;/p&gt;

&lt;p&gt;The question the solicitation leaves behind is not whether the agent attacks well, I ask myself: what can be proven afterwards?&lt;/p&gt;

&lt;p&gt;And that question is not theoretical for defenders, because without an artifact there is no forensics and without forensics there is no attribution.&lt;/p&gt;

&lt;p&gt;There is one design detail worth looking at slowly, and it is the copy to other machines.&lt;/p&gt;

&lt;p&gt;An agent that replicates does not need an operator behind every machine, it needs a channel and an excuse to use it.&lt;/p&gt;

&lt;p&gt;That channel is what remains when the agent is gone.&lt;/p&gt;

&lt;h2&gt;
  
  
  The other extreme
&lt;/h2&gt;

&lt;p&gt;The same day the solicitation became known, Microsoft expanded Project Perception with Guardian Agents (Source: helpnetsecurity.com).&lt;/p&gt;

&lt;p&gt;What it sells is the opposite, curiously: its own identity for each agent in Entra Agent ID, unified logging and containment when the agent goes off script.&lt;/p&gt;

&lt;p&gt;The tool already discovers 60 types of harness, including Claude Code and Codex.&lt;/p&gt;

&lt;p&gt;That is, the market is moving in two directions at once.&lt;/p&gt;

&lt;p&gt;1) One side buys the agent that disappears without a trace.&lt;/p&gt;

&lt;p&gt;2) The other sells the agent that leaves identity, logs and a way to stop it.&lt;/p&gt;

&lt;p&gt;Anyone who has to defend a network will have to choose which side of that scale they work on.&lt;/p&gt;

&lt;h2&gt;
  
  
  The other side
&lt;/h2&gt;

&lt;p&gt;Agents did not invent erasing the trail.&lt;/p&gt;

&lt;p&gt;This same week, AOMEI suffered the compromise of its servers to distribute a backdoor through the promotional feature built into Backupper (Source: bleepingcomputer.com).&lt;/p&gt;

&lt;p&gt;The detail is the interesting part: the attackers restored the configuration files after every push to erase the trail.&lt;/p&gt;

&lt;p&gt;Symantec attributes to APT28 a campaign with the Hookedge backdoor against a naval shipyard, a military aviation engine manufacturer and a police agency in two NATO countries (Source: therecord.media).&lt;/p&gt;

&lt;p&gt;They got in with a malicious Word document in July and stayed active until September, that is, three months inside an industrial network without anyone seeing it.&lt;/p&gt;

&lt;p&gt;The difference with ECHOTRIBBLE is one of degree, not of substance, because they want to go unnoticed and the agent is designed to disappear.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who answers?
&lt;/h2&gt;

&lt;p&gt;There is the gap nobody has closed.&lt;/p&gt;

&lt;p&gt;Dark Reading raised it this week: if an agent causes a breach, the technical responsibility is human, but the legal one is still unassigned (Source: darkreading.com).&lt;/p&gt;

&lt;p&gt;OpenAI and Anthropic would not comment on whether the model operator commits the breach.&lt;/p&gt;

&lt;p&gt;With an agent that self-destructs, the question gets complicated, honestly.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;If there is no record of the agent, what do you charge someone with?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The easy answer is to blame whoever signed off on the operation. The useful one, I would say, is that the operation leaves proof before it runs.&lt;/p&gt;

&lt;p&gt;And that is decided in the design, not at trial.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to look at
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Endpoint logging.&lt;/strong&gt; If the agent lives only in memory and leaves no process event, there is nothing to investigate. Check which telemetry survives the self-destruction.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The copy.&lt;/strong&gt; An agent that replicates to other machines needs a channel. That channel, whether SMB, a share or device management, is where it shows.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Oversight.&lt;/strong&gt; "Minimal human oversight" means someone signs off on the operation. Ask who signs and with what information.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The vendor side.&lt;/strong&gt; If your agent tool has no per-agent identity, its own logging and a way to revoke, it is not that the attacker will not use it. It is that you cannot audit it either.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How would I test it in my lab?
&lt;/h2&gt;

&lt;p&gt;I would set up a local agent on a VM with a narrow task and a log of every tool call with its arguments, and I would ask it to copy itself to a second VM.&lt;/p&gt;

&lt;p&gt;Then I would delete the agent from the first one and look at what is left in the SIEM.&lt;/p&gt;

&lt;p&gt;What I am after is not whether the copy works, but how much survives the deletion.&lt;/p&gt;

&lt;p&gt;I would repeat the test with the log outside the compromised machine, on a separate collector, to see if the self-destruction takes it down too.&lt;/p&gt;

&lt;p&gt;If the log does not capture it, the problem is not the agent: it is the instrumentation.&lt;/p&gt;

&lt;h2&gt;
  
  
  Closing
&lt;/h2&gt;

&lt;p&gt;ECHOTRIBBLE does not show a new agent: it shows that the trail has become the battlefield.&lt;/p&gt;

&lt;p&gt;I know I keep going on about it, but this "trail" thing is something I have been talking about in every post I write, control and audit.&lt;/p&gt;

&lt;p&gt;A defender does not need to see the agent in real time, or at least not constantly, but they do need the agent to be unable to leave without a record that it was there.&lt;/p&gt;




&lt;p&gt;Originally published at &lt;a href="https://sammideblas.com/notas/us-army-is-buying-agents-that-self-destruct-before-detection" rel="noopener noreferrer"&gt;https://sammideblas.com/notas/us-army-is-buying-agents-that-self-destruct-before-detection&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Si has leído hasta aquí... reacciona y comparte -&amp;gt; "la seguridad y la defensa en la era de la IA es cosa de todos los que la usan" ¿No Crees?&lt;/p&gt;

</description>
      <category>ai</category>
      <category>security</category>
      <category>agents</category>
      <category>logging</category>
    </item>
  </channel>
</rss>
