---
title: "Oct 20–26: LiteLLM poisoned to steal AI keys, and more"
show: "RobotPaper's weekly AI security news"
canonical: "https://podcasts.robotpaper.ai/sec/ep/2026-10-26-week/"
date: "2026-10-26T07:08:53+00:00"
duration: "27:10"
duration_seconds: 1630
audio: "https://podcasts.robotpaper.ai/sec/ep/2026-10-26-week/2026-10-26-week-sec.mp3"
chapters: "https://podcasts.robotpaper.ai/sec/ep/2026-10-26-week/chapters.json"
transcript: "https://podcasts.robotpaper.ai/sec/ep/2026-10-26-week.txt"
feed: "https://podcasts.robotpaper.ai/sec/feed.xml"
language: "en"
author: "Roy Osherove"
ai_generated_voices: true
---

# Oct 20–26: LiteLLM poisoned to steal AI keys, and more

RobotPaper's weekly AI security news · Oct 26, 2026 · 27 min

- Listen: [MP3, 26 MB](https://podcasts.robotpaper.ai/sec/ep/2026-10-26-week/2026-10-26-week-sec.mp3)
- Subscribe via RSS: https://podcasts.robotpaper.ai/sec/feed.xml (paste into any podcast app)
- Web page: https://podcasts.robotpaper.ai/sec/ep/2026-10-26-week/
- Chapters (JSON): https://podcasts.robotpaper.ai/sec/ep/2026-10-26-week/chapters.json
- Transcript (text): https://podcasts.robotpaper.ai/sec/ep/2026-10-26-week.txt
- Show: [RobotPaper's weekly AI security news](https://podcasts.robotpaper.ai/sec/) (RSS: https://podcasts.robotpaper.ai/sec/feed.xml)

## Chapters

- `00:00:00` Intro
- `00:00:10` LiteLLM poisoned
- `00:02:20` AI key custodians hit
- `00:04:26` AGENTS.md hijack
- `00:06:19` MCP servers run code
- `00:08:24` Malware advisories
- `00:10:14` OVHcloud CLI swap
- `00:11:58` PHP2Shell
- `00:13:58` eMudhra CA keys
- `00:16:06` Brotli decoder flaw
- `00:17:45` Plex2Shell
- `00:18:46` AWS eMudhra bulletin
- `00:20:09` Agent bulletins
- `00:21:22` AI-found browser bugs
- `00:22:42` Abliterated models
- `00:23:51` GitLab and Arista
- `00:25:06` ImageMagick and BlackX
- `00:26:04` DreamHost incident
- `00:26:28` Outro

## Show notes

### AI infrastructure and API keys (4)

- [The attack chain behind the v1.107.0 LiteLLM release](https://docs.litellm.ai/blog/security-incident-oct-21-2026-attack-chain) · [also: Microsoft Security Blog](https://www.microsoft.com/en-us/security/blog/2026/10/23/inside-the-litellm-1-107-0-supply-chain-compromise-credential-theft-through-python-startup-hooks/) · [also: BleepingComputer](https://www.bleepingcomputer.com/news/security/litellm-release-pipeline-hacked-to-distribute-info-stealer-malware/)
- [Hackers hit AI coding tool OpenCode, running up charges on users’ AI accounts](https://techcrunch.com/2026/10/19/hackers-hit-ai-coding-tool-opencode-running-up-charges-on-users-ai-accounts/) · [also: Railway](https://blog.railway.com/p/security-disclosure-wednesday-october-21st) · [also: Railway](https://blog.railway.com/p/preliminary-report-october-security-incident) · [also: Railway Status](https://status.railway.com/incident/8E4V5WLD) · [also: Zeabur](https://zeabur.com/blogs/october-2026-security-incident)
- [AI coding agents hijacked using AGENTS.md file in Downloads folder](https://www.bleepingcomputer.com/news/security/ai-coding-agents-hijacked-using-agentsmd-file-in-downloads-folder/)
- [Open and Shut: Remote Code Execution in opencode via MCP OAuth](https://www.pillar.security/blog/open-and-shut-remote-code-execution-in-opencode-via-mcp-oauth) · [also: SpanDock](https://www.spandock.com/blog/mcp-server-entries-are-commands) · [also: GitHub Advisory Database](https://github.com/advisories/GHSA-h4mw-j7qp-8mwm)

### Package registries and release integrity (2)

- [GitHub Advisory Database: malware advisories](https://github.com/advisories?query=type%3Amalware) · [also: GitHub Advisory Database](https://github.com/advisories/GHSA-jgm6-7jxr-97gx) · [also: GitHub Advisory Database](https://github.com/advisories/GHSA-mvgr-2gjf-h99c) · [also: GitHub Advisory Database](https://github.com/advisories/GHSA-2hr3-wchh-6g2q) · [also: GitHub Advisory Database](https://github.com/advisories/GHSA-g5mh-f2v3-6p25)
- [OVHcloud CLI Compromised via GitHub Release Swap: Detonating the Real GlassWorm Loader and What We Found](https://www.stepsecurity.io/blog/ovhcloud-cli-github-release-swap-glassworm)

### Critical vulnerabilities in widely deployed software (4)

- [Today's PHP Security Release, and How We're Strengthening PHP Security](https://thephp.foundation/blog/2026/10/21/todays-php-security-release-and-how-were-strengthening-php-security/) · [also: BleepingComputer](https://www.bleepingcomputer.com/news/security/amazon-says-php2shell-exploitation-began-4-days-before-patches/)
- [Protecting Chrome and Android Users from Compromised eMudhra Intermediate CA Certificates](https://blog.google/security/protecting-chrome-and-android-users-from-compromised-emudhra-intermediate-ca-certificates/) · [also: Aikido](https://www.aikido.dev/blog/emudhra-ca-compromise) · [also: Python.org](https://www.python.org/downloads/release/python-31318/)
- [Wednesday, October 21, 2026 Security Releases](https://nodejs.org/en/blog/vulnerability/october-2026-security-releases) · [also: Aikido](https://www.aikido.dev/blog/python-packages-vulnerable-node2shell-brotli-cve-2026-111497) · [also: Ubuntu](https://ubuntu.com/security/CVE-2026-111497)
- [PSA: Patch your Plex media server before someone hacks it](https://arstechnica.com/security/2026/10/recently-fixed-plex-media-vulnerabilities-are-being-exploited-to-install-malware/) · [also: CISA](https://www.cisa.gov/news-events/alerts/2026/10/20/cisa-adds-six-known-exploited-vulnerabilities-catalog)

### Amazon Web Services security bulletins (2)

- [CVE-2026-111943: eMudhra intermediate CA key recovery in AWS trust bundles](https://aws.amazon.com/security/security-bulletins/2026-150-aws/)
- [CVE-2026-115845 - Incomplete validation in the run\_script Python sandbox in AWS MCP Server](https://aws.amazon.com/security/security-bulletins/2026-149-aws/) · [also: AWS security bulletin](https://aws.amazon.com/security/security-bulletins/2026-146-aws/)

### AI and vulnerability research (2)

- [Stable Channel Update for Desktop](https://chromereleases.googleblog.com/2026/10/stable-channel-update-for-desktop_02018666648.html) · [also: SecurityWeek](https://www.securityweek.com/memory-bugs-account-for-most-of-chromes-353-new-security-fixes/) · [also: The PHP Foundation](https://thephp.foundation/blog/2026/10/21/todays-php-security-release-and-how-were-strengthening-php-security/)
- [What Security Teams Need to Know About Agents Using Abliterated Models](https://eve.security/blog/what-security-teams-need-to-know-about-agents-using-abliterated-models) · [also: Railway](https://blog.railway.com/p/preliminary-report-october-security-incident) · [also: TechCrunch](https://techcrunch.com/2026/10/19/hackers-hit-ai-coding-tool-opencode-running-up-charges-on-users-ai-accounts/)

### Quick hits (3)

- [GitLab warns of new max severity code execution vulnerability](https://www.bleepingcomputer.com/news/security/gitlab-warns-of-new-max-severity-code-execution-vulnerability/) · [also: BleepingComputer](https://www.bleepingcomputer.com/news/security/arista-patches-new-velocloud-orchestrator-zero-day-exploited-in-attacks/) · [also: BleepingComputer](https://www.bleepingcomputer.com/news/security/hpe-patches-eight-max-severity-rce-and-auth-bypass-aos-flaws/)
- [CVE-2026-116149](https://security-tracker.debian.org/tracker/CVE-2026-116149) · [also: BleepingComputer](https://www.bleepingcomputer.com/news/security/blackx-uefi-bootkit-exploits-baton-drop-flaw-to-bypass-secure-boot/)
- [Investigating an incident affecting File Manager and Webmail](https://www.dreamhoststatus.com/)

The voices you're hearing are AI-generated; the stories come from public sources linked in the show notes.\
Music: Eleven Music.\
An independent project; not affiliated with or endorsed by Amazon Web Services, Y Combinator, Hacker News or any vendor or news outlet mentioned. Opinions are the hosts' own.\
Produced by Roy Osherove.

## Transcript

\[Linus\] This is RobotPaper's weekly AI security news; the voices you're hearing are AI-generated; the stories come from public sources.

\[Grace\] AI infrastructure and A P I keys. This is the security news for October nineteenth through October twenty-sixth, and it opens with attacks on the places that hold the keys to AI services.

LiteLLM is an open-source gateway that sits between applications and AI model providers, so one service can route requests to many models. Because of that position, it can hold provider A P I keys and database credentials. On October twenty-first, a compromised version, one point one-oh-seven, was published to the Python Package Index, known as Py P I, and to container registries. This is a supply-chain attack: instead of attacking victims directly, the attacker poisons software they already trust and install.

Microsoft's security team describes how it worked. The package placed startup files in Python's library directory, so the malicious code could run whenever Python started, even if LiteLLM was never imported. It collected credentials for cloud accounts, developer and AI tools, Git, S S H keys and browsers, and sent them out encrypted. Microsoft's guidance includes not running Python or pip inside a suspect environment, because that can trigger the code again.

LiteLLM's own root cause analysis says the chain began in May, when a master key for an experimental internal tool was committed to a public repository and never revoked. On October tenth, an attacker used it to obtain a GitHub token with organization-owner rights, stole more tokens through temporary workflows, pointed the release process at a lookalike repository, and approved the release with another stolen token.

The Py P I release was quarantined in under two hours, and the container tags were cleaned up about two and a half hours after publication. If you installed that version, or pulled the latest, main-latest or main-stable images that day, treat the machine as compromised. Rebuild from clean artifacts, and rotate every reachable secret from a clean machine. Pin packages by hash and containers by digest, not by moving tags.

\[Ken\] Next up, a related pattern: platforms that store customers' AI provider keys were hit within days of each other. An A P I key is a secret string that lets software call a service, so anyone who holds it can run up charges on the owner's account.

OpenCode, an open-source AI coding agent, said on Saturday, October seventeenth, that attackers broke into an older system behind its hosted Console and affected one point four percent of Console users. The company called it a large-scale distillation attack; distillation means using a stronger model's outputs to train another model. TechCrunch reports that users described unauthorized charges in the hundreds of dollars, with a few above one thousand, and that some skeptics say the incident was a plain breach whose use for distillation is unproven.

Railway, a cloud hosting platform, reported that an intruder escalated privileges in its build infrastructure between October twentieth and twenty-second, and decrypted environment variables for some customers. Railway says at most one point eight five percent of workspaces, roughly a hundred thousand, had variables potentially exposed. It asked providers including Anthropic and OpenAI to revoke keys, paused builds, and says running services stayed online. A notice on October twenty-fifth adds that later revocations were precautionary, and that some replacement keys were revoked by mistake.

Zeabur, another hosting platform, said in a preliminary report that an attacker used leaked internal credentials on October fourteenth and fifteenth to read environment variables from its database. Zeabur said the attacker appeared to be searching mainly for AI service keys, and it has announced it will shut down on December thirty-first.

Railway describes its incident as part of a series of attacks on platform companies. The reports covered here do not name a common attacker.

If you stored provider keys on any of these platforms, rotate them at the provider that issued them. Deleting a key from a dashboard does not revoke it. Check provider billing since early October, use a separate key for each tool, and set spend alerts.

\[Dennis\] AGENTS dot M D and CLAUDE dot M D are plain-text instruction files that AI coding agents, such as OpenAI Codex and Claude Code, load automatically from their working directory. They tell the agent how to build, test and behave in a project. BleepingComputer reports that researchers at Forever Security found those files can be turned into an attack.

They discovered it on a customer's computer, where a coding agent tried to install a malicious N P M package named function-color. The trail led to an AGENTS dot M D file in the user's Downloads folder, which was also the agent's working directory. The researchers say a web page can silently download a file with that name into Downloads, because many browsers save files there without asking. BleepingComputer's own test found Chrome, Edge and Firefox did this, while Brave prompted the user to approve the download.

The attack then needs the person to start an AI agent in that folder, which the researchers say is common, for example to organize downloads. Forever Security says Claude Code alone launched more than thirteen hundred sessions from Downloads at its customers. In their tests against Codex, a harmless payload ran ninety percent of the time. A malicious CLAUDE dot M D redirected Claude Code's A P I traffic to a server the researchers controlled, which exposed the agent's O Auth token and led to code execution in roughly half of tests.

The researchers say there is nothing for OpenAI or Anthropic to patch, because the agent is doing what it was built to do. The advice: start coding agents only in the specific project directory, never in Downloads or your home folder. Review unfamiliar instruction files before an agent loads them, and keep approval prompts and sandboxing on.

\[Ada\] The Model Context Protocol, or M C P, is a standard way for AI agents to connect to outside tools and data sources. Two analyses this week show how a connection can turn into code execution.

Pillar Security disclosed a critical flaw in opencode, which it says has about three million weekly N P M downloads. On Windows and the Windows Subsystem for Linux, a malicious remote M C P server could run code on a developer's machine during the O Auth login step, with no extra click. Pillar stresses that it needs no prompt injection, which means hiding hostile instructions in text a model processes. The cause was that a login address from the server was passed to a PowerShell command. The fix is in opencode version one point eighteen point thirty-three, released September twenty-eighth. Pillar says the vendor shipped it as a refactor, with no advisory and no C V E, so scanners will not flag old installs. A C V E is a Common Vulnerabilities and Exposures identifier, a public number for one specific flaw.

A separate analysis from SpanDock groups three October records: JetBrains Junie C L I, fixed in version twenty-six point eight point three; Z dot ai's ZCode, where the record names no fixed version; and the Claude Agent S D K for Python. Anthropic's advisory for that one rates it nine point two on C V S S version four. C V S S is a zero to ten severity score, where ten is the most severe. The common thread is that a local M C P server entry is simply a command the tool starts. SpanDock notes that interactive tools usually ask before loading a project's servers, but the vendors' own documents say headless runs, such as C I jobs, skip that prompt.

If you use these tools, update them, including the Python agent S D K to version zero point two point one-two-four or later. Treat M C P settings in cloned repositories as executable, and connect only to servers you trust.

\[Linus\] Package registries and release integrity. GitHub's advisory database added sixty-four malware advisories between October twenty-fourth and twenty-sixth: sixty for N P M packages and four for Py P I. A malware advisory is an entry that flags a published package as malicious, so tools and developers can find and remove it.

Most of the N P M entries are stub packages published under very high version numbers, such as ninety-nine point nine point one. That is the pattern of dependency confusion, where an attacker publishes a public package with the same name as a company's internal one, hoping a build tool picks the higher version. The advisories, which cite Amazon Inspector and the OpenSSF Package Analysis project, describe many of these as sending host details back to the publisher when installed.

The four Py P I packages have names that look like AI tooling. The advisory for claudeopusfile describes obfuscated install code that fetches a hidden payload, an infostealer whose control server address is found in blockchain transaction history. For nebulallm, it describes a typosquat whose install script downloads a script and then a malicious D L L. The advisories for zagents-message-bus and fontus describe limited risk: they are categorized as probable penetration-test packages that collect basic host information.

The practical point is that coding agents install packages on their own. Check lockfiles and virtual environments for these names, consider an internal registry mirror or a waiting period for brand-new packages, and review any unfamiliar package an agent adds.

\[Grace\] The OVHcloud command-line tool was compromised through its GitHub releases, according to a StepSecurity analysis. On October nineteenth, between twelve forty-seven and fourteen twenty-one U T C, an attacker using an OVHcloud contributor's GitHub account replaced the binaries on five published releases, version zero point eleven through zero point fifteen, with a credential stealer. No tag was moved and the source code was untouched. The checksum file was replaced last, so the published hashes matched the swapped files. A checksum is a fingerprint of a file, used to confirm it has not been altered.

The swap lasted ninety-four minutes, with about two hundred fifty downloads, and a customer caught it because their C I pinned the S H A two fifty-six hash of one download. StepSecurity says the root cause was that GitHub release immutability was not enabled. It reports the loader belongs to the GlassWorm family, and that the stealer's targets include AI-agent credential stores, cloud command-line tools, Docker and Kubernetes configuration, N P M tokens and S S H keys. OVHcloud's postmortem traces initial access to a malicious VS Code extension installed on a contributor's personal laptop a year earlier.

If you downloaded the tool from GitHub releases in that window, treat the host as compromised, rotate secrets from a different machine, and look for temporary directories whose names start with envchecker. Homebrew, go install, source builds and the Docker image were not exposed. Running the tool's version command on a swapped build printed undefined. The broader lesson is to pin hashes of downloaded release binaries.

\[Ken\] Critical vulnerabilities in widely deployed software. P H P is a widely used server-side web language; BleepingComputer says it runs up to eighty percent of websites, and VirusTotal tracks more than five million exposed P H P hosts.

On October twenty-first, the P H P project released versions eight point five point twelve, eight point four point twenty-seven, eight point three point thirty-six and eight point two point thirty-five. They fix two critical flaws nicknamed P H P two Shell, which allow remote code execution: an attacker running their own commands on a server over the network. According to the P H P Foundation, one of them, tracked as C V E twenty twenty-six, one zero nine six one eight, involves memory corruption when a request hits the memory limit, and on Apache with mod P H P an unauthenticated request can trigger it before any application code runs. The other, one zero nine six one nine, involves leftover state from a previous request. The Foundation says the score is nine point two, not the ten that circulated in a leak.

That leak forced the release forward a day. A security firm also mistakenly published a proof-of-concept video, and a distributor published fixed container images hours early. Amazon Threat Intelligence says exploitation attempts began October seventeenth, four days before the patch, and that the flaws are now broadly exploited. Other firms saw only failed attempts, and the P H P team said none of the early public exploit code worked at release. Dataminr reported no known compromises as of October twenty-second.

E P S S, the Exploit Prediction Scoring System, estimates the chance a flaw will be exploited soon; both scores are near the top of the scale, about ninety-nine point nine eight percent. If you run P H P, update to a fixed release now. Unsupported branches get no official fix, so use your distribution's backport or upgrade.

\[Dennis\] A certificate authority, or C A, is an organization that browsers and operating systems trust to vouch that a website's certificate really belongs to it. Intermediate C A certificates are issued by a root and can then sign website certificates.

On October seventeenth, Google's Chrome Root Program received a report that the private keys for eight eMudhra intermediate C A certificates were generated with too little randomness. Anyone could forge trusted H T T P S certificates for any site, against clients that still trusted those C As. Chrome blocked the keys within two hours using a CRL set, its list of revoked certificates. Google notes that no certificates from these C As were logged in Certificate Transparency, the public logs of issued certificates, so clients that enforce that, including Chrome, could not have been attacked. Aikido reports the weakness was extreme: one key is just the number four.

Chrome is also removing the emSign E C C Root C A G three on Tuesday, October twenty-seventh. If your certificates chain to it, get a new one from another C A now.

The fallout reached tools with their own trust lists. Node.js updated its bundled certificates on October twenty-first. On October twenty-fifth, Python released three point thirteen point eighteen, three point twelve point sixteen and three point eleven point eighteen, adding a denylist of the compromised keys to its S S L module. That is C V E twenty twenty-six, one one five five five seven, scored nine point four by Python's own C V E assigner. The same releases fix a flaw in the tarfile module's extraction filter, C V E twenty twenty-six, one one four five eight three, where crafted links could write files outside the destination directory.

Update your operating system's C A certificates, rebuild container images, and move to the fixed Node and Python releases.

\[Ada\] Brotli is a compression format from Google, widely used for web content. C V E twenty twenty-six, one one one four nine seven is an out-of-bounds write in its decoder, in versions one point zero point three through one point two point zero. Malformed compressed data can crash a program, leak memory or let an attacker run code. Google's C V E record scores it nine point two.

Node.js fixed it on October twenty-first, in versions twenty-two point twenty-three point four, twenty-four point twenty-two point one and twenty-six point twelve point one, crediting OpenAI Codex Security. Aikido reports it has since spread: the Python Brotli and brotlicffi packages are affected, fixed in one point two point one, as are Bun, Envoy and Istio, and curl-based tools. It works in both directions, so a malicious server can attack an unpatched client. Aikido adds that OpenAI found the root cause in February, but it was not recognized as a security flaw and no release carried the fix until now. Apache's mod brotli only compresses and is not affected.

The E P S S score is zero point three three, in the top two percent. As of October twenty-sixth, Ubuntu's tracker still lists the brotli package as needing evaluation, with no fix.

Upgrade Node and rebuild container images, because an image pinned to an older version stays vulnerable. If you cannot patch, reject requests that use Brotli encoding and stop advertising it in outgoing requests.

\[Linus\] Plex Media Server is software for streaming a personal media library, and it is often exposed to the internet. Ars Technica reports that a researcher chained four recently fixed Plex flaws into a remote code execution attack called Plex two Shell, using AI heavily to develop it. The chain includes S S R F, server-side request forgery, which tricks a server into making requests on an attacker's behalf. A second researcher's honeypot saw active attacks installing cryptominers and residential proxy software.

On October twentieth, CISA added the four Plex flaws to its Known Exploited Vulnerabilities catalog, or K E V, a list of flaws with evidence of active exploitation that U S federal agencies must fix by set deadlines. Update to Plex version one point forty-three point three or later, and keep the server off the open internet or behind a reverse proxy.

\[Grace\] Amazon Web Services security bulletins. AWS bulletin twenty twenty-six dash one fifty covers the eMudhra problem. It says eight intermediate keys, six for T L S and two for code signing, were trivially recoverable, and that the issuing C A revoked them on October eighteenth and nineteenth. Customers whose T L S clients enforce revocation checking are not affected.

AWS removed three affected roots from its shared trust store on October twenty-first and released boto3 and botocore one point forty-three point one seventeen. Updates for Amazon Linux and Amazon Corretto, Amazon's Java distribution, are available. Amazon CloudFront, a content delivery network, is removing the roots from its edge bundle by October twenty-eighth. Elastic Load Balancing, which spreads traffic across servers, is updating Application Load Balancers with O I D C or J W T authentication between October twenty-ninth and thirty-first, with affected customers told through the AWS Health dashboard. If your identity provider's certificate chains to an affected root, replace it by October twenty-eighth. Bottlerocket, Amazon's container-focused operating system, users on older images should apply the interim trust override from the bulletin until new images arrive.

\[Ken\] Two bulletins concern agent tooling. The AWS M C P Server is a managed server that gives AI agents access to AWS through M C P, including a run script tool for sandboxed Python. Bulletin one forty-nine says its code validator had an incomplete list of disallowed inputs, so an authenticated user could run code inside their own isolated session. AWS fixed it on the service side, and no customer action is needed.

Bulletin one forty-six covers the Strands Agents S D K, an open-source toolkit for building AI agents. Its agent-to-agent server exposes an agent over H T T P, and it did not isolate conversation state between callers. In multi-tenant deployments, one authenticated user could see another caller's conversation history or influence the model's behavior. It is fixed in strands-agents one point forty-three for Python and in version one point five for the TypeScript package. There is no workaround; until you upgrade, do not expose one server to several tenants.

One caution: older bulletins from early September reappeared in the bulletin feed with October twenty-fifth timestamps. They are not new.

\[Dennis\] AI and vulnerability research. Chrome one fifty-six reached the stable channel on October twentieth with fixes for three hundred fifty-three security issues, according to SecurityWeek. Six are critical and sixty-one high, and most are memory-safety bugs such as use-after-free flaws, where a program keeps using memory after releasing it. Google reports no known exploitation.

The credits show who is finding bugs. No bounty was paid on two hundred eighty-five of the issues: Google's own team found two hundred sixty-six, and the rest came from a researcher listed as working for Anthropic and assisted by Claude, and from OpenAI Codex Security.

That fits other reporting this week. The P H P Foundation credits OpenAI Codex Security and the firm Wiz for work on the P H P flaws, and says AI-assisted research has increased the number of reports it receives. It has published fifteen C V Es since July first, and handled more than a hundred reports that were invalid, duplicates or not reproducible. AI helps its team validate and group reports, but a person makes each call. Update Chrome and restart it.

\[Ada\] Eve Security, a vendor, published a post on October twelfth, which reached the Hacker News front page on October twenty-fifth, about agents that use abliterated models. Abliteration removes an open-weight model's refusal behavior by modifying its weights or activations. The post argues that model safeguards do not bind someone who runs their own copy, so controls must sit at the agent and network layer. It says Anthropic's testing showed an abliterated model complying with harmful cyber-attack instructions in every case. This is a vendor opinion piece, not an incident report.

Other reporting adds context. Railway says early evidence suggests its attacker had access to abliterated models. TechCrunch reports OpenCode's team found AI models refusing to analyze malicious code while defending against its breach, and that it later gained access to an OpenAI program with fewer refusals for defensive security work.

If you run agents, rely on scoped credentials, restricted outbound access and audit logs rather than on model refusals.

\[Linus\] Quick hits. GitLab patched C V E twenty twenty-six, one one zero two six four in several release lines, including nineteen point five point one. BleepingComputer says an unauthenticated user could run code on the server through an integer overflow in a regular expression compiler, and that Shadowserver counts more than twenty-seven hundred exposed community edition instances. It is the second maximum-severity flaw of that kind in about ten days. If you self-host GitLab, upgrade.

Arista patched an actively exploited zero-day, C V E twenty twenty-six, one one four six two five, in its on-premises VeloCloud Orchestrator. A zero-day is a flaw exploited before a patch exists. Any authenticated user could become a superuser operator through S Q L injection. CISA added it to K E V on October twenty-third. Restrict the management interface to administrator networks. BleepingComputer also reports that H P E patched eight maximum-severity remote code execution and authentication bypass flaws in its ArubaOS software.

\[Grace\] ImageMagick received twenty-one new C V Es on October twenty-fifth. The highest, scored seven point two, is a stack buffer overflow when decoding bzip2-compressed images. Upstream fixes are in seven point one point two dash thirty-three and six point nine point thirteen dash fifty-eight. Debian's tracker still lists bookworm and trixie as vulnerable, with fixes only in unstable. If you process untrusted images, restrict formats in the policy file and run it sandboxed.

BlackX is a U E F I bootkit, which infects the earliest startup code. Kaspersky research, reported by BleepingComputer, links it with medium confidence to APT forty-one. It bypasses Secure Boot, a feature that only lets trusted code start the operating system, using the old Baton Drop flaw. All victims ran fully patched Windows, because the fix requires a manual certificate revocation that does not happen automatically.

\[Ken\] The weekend was quiet. CISA added nothing to the K E V catalog after October twenty-third. DreamHost reported an incident affecting its File Manager and webmail, and reset passwords for some mail and F T P users. Account owners must set new passwords and share them. DreamHost says it will never email you asking for your password, and the cause has not been stated.

\[Linus\] That concludes RobotPaper's weekly AI security news. The week's theme was attackers targeting the places that hold AI keys and agent credentials, while old trust assumptions about certificates, release files and working directories gave way. The practical habits repeat: rotate keys at the issuer, pin hashes and digests, run agents only in project folders, and keep trust stores current. From Linus, Dennis, Ken, Ada, and Grace, thank you for listening.

---

The voices you're hearing are AI-generated; the stories come from public sources linked in the show notes. An independent project; not affiliated with or endorsed by Amazon Web Services, Y Combinator, Hacker News or any vendor or news outlet mentioned. Opinions are the hosts' own. Produced by Roy Osherove. Privacy: We count visits to these pages with Cloudflare Web Analytics (cookieless). Google Analytics (which sets cookies): visitors in the EU/EEA, the UK and Switzerland are asked first and it runs only if they accept (we go by the country Cloudflare reports for the connection; if it is unknown, we ask). Elsewhere it loads by default. Anyone can turn it off with the Privacy settings button at the bottom of every page; the choice is kept for a year in this browser (local storage "analytics-consent"). We count feed fetches and downloads in aggregate; IP addresses are hashed and never stored. If you send feedback, we keep what you write (and your email if you give it) to read and reply. © 2026 Roy Osherove.
