---
title: "Nov 10–16: Cursor breach ships trojanized extension, and more"
show: "RobotPaper's weekly AI security news"
canonical: "https://podcasts.robotpaper.ai/sec/ep/2026-11-16-week/"
date: "2026-11-16T03:17:17+00:00"
duration: "29:29"
duration_seconds: 1769
audio: "https://podcasts.robotpaper.ai/sec/ep/2026-11-16-week/2026-11-16-week-sec.mp3"
chapters: "https://podcasts.robotpaper.ai/sec/ep/2026-11-16-week/chapters.json"
transcript: "https://podcasts.robotpaper.ai/sec/ep/2026-11-16-week.txt"
feed: "https://podcasts.robotpaper.ai/sec/feed.xml"
language: "en"
author: "Roy Osherove"
ai_generated_voices: true
---

# Nov 10–16: Cursor breach ships trojanized extension, and more

RobotPaper's weekly AI security news · Nov 16, 2026 · 29 min

- Listen: [MP3, 28 MB](https://podcasts.robotpaper.ai/sec/ep/2026-11-16-week/2026-11-16-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/2026-11-16-week/
- Chapters (JSON): https://podcasts.robotpaper.ai/sec/ep/2026-11-16-week/chapters.json
- Transcript (text): https://podcasts.robotpaper.ai/sec/ep/2026-11-16-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` Cursor breach
- `00:03:04` Fireworks hits Notion
- `00:04:19` LiteLLM 1.107.0
- `00:06:30` FBI ARTEX warning
- `00:08:50` Glasswing ledger
- `00:10:37` AWS MCP flaws
- `00:12:10` Apache HTTP/2 RCE
- `00:14:11` React Flight RCE
- `00:15:31` Patch Tuesday
- `00:17:27` GPUStack auth bypass
- `00:18:26` Barracuda KEV
- `00:19:28` Siemens advisories
- `00:20:36` sshole lab test
- `00:21:36` PyPI provenance
- `00:23:10` Binance dev machine
- `00:24:10` AWS Go wrapper, pandas
- `00:25:17` Malware advisories
- `00:26:21` Project Zero 90 days
- `00:27:29` Humble Bundle login
- `00:28:04` Postgres, Caixa
- `00:28:42` Outro

## Show notes

### AI and agent security (6)

- [Cursor confirms security breach led to malicious Claude Code extension](https://www.bleepingcomputer.com/news/security/cursor-confirms-security-breach-led-to-malicious-claude-code-extension/) · [also: StepSecurity](https://www.stepsecurity.io/blog/claude-code-extension-trojanized)
- [Fireworks AI security breach impacts Notion and their customers](https://www.notion.com/help/security-incident-involving-a-third-party-ai-vendor)
- [litellm 1.107.0 Compromised on PyPI and Docker: Credential Stealer and Hourly Remote Code Loader](https://www.stepsecurity.io/blog/litellm-1-107-0-compromised-pypi-backdoored-wheels-credential-stealer)
- [FBI warns ARTEX AI hacking tool is powering global cyberattacks](https://www.bleepingcomputer.com/news/security/fbi-warns-artex-ai-hacking-tool-is-powering-global-cyberattacks/) · [also: StepSecurity](https://www.stepsecurity.io/blog/artex-ai-pentesting-tool-attackers-leaked-secrets)
- [Tracking Anthropic-discovered vulnerabilities](https://www.vulncheck.com/blog/anthropic-glasswing-november-update)
- [CVE-2026-125271 - AWS API MCP Server Security Policy Bypass via Incorrect Operation Name Normalization](https://aws.amazon.com/security/security-bulletins/rss/2026-176-aws/) · [also: AWS security bulletin](https://aws.amazon.com/security/security-bulletins/rss/2026-180-aws/) · [also: AWS security bulletin](https://aws.amazon.com/security/security-bulletins/rss/2026-179-aws/)

### Critical vulnerabilities in widely deployed software (7)

- [Critical Apache HTTP Server vulnerability can allow remote code execution (CVE-2026-119618)](https://www.aikido.dev/blog/cve-2026-119618-apache-http-server-rce) · [also: httpd.apache.org](https://httpd.apache.org/security/vulnerabilities_24.html)
- [CVE-2026-126187: React Server Components Flaw Lets Attackers Execute Code Through Smuggled Flight Rows](https://www.stepsecurity.io/blog/cve-2026-126187-react-flight-rce) · [also: GitHub Advisory Database](https://github.com/advisories/GHSA-xqc2-xf4c-x3h3)
- [Microsoft Fixes Nearly 1,100 Security Vulnerabilities](https://krebsonsecurity.com/2026/11/microsoft-fixes-nearly-1100-security-vulnerabilities/) · [also: CISA KEV](https://nvd.nist.gov/vuln/detail/CVE-2026-124194) · [also: CISA KEV](https://nvd.nist.gov/vuln/detail/CVE-2026-125350)
- [Exploit Attempts for GPUStack Authentication Bypass (CVE-2026-102895), (Thu, Nov 12th)](https://isc.sans.edu/diary/rss/33514)
- [CVE-2026-111560: Barracuda Networks Multiple Products OS Command Injection Vulnerability](https://nvd.nist.gov/vuln/detail/CVE-2026-111560) · [also: CISA KEV](https://nvd.nist.gov/vuln/detail/CVE-2026-116606)
- [CISA ICS advisory: Siemens Desigo, Gridscale](https://www.cisa.gov/news-events/ics-advisories/icsa-26-316-02) · [also: CISA](https://www.cisa.gov/news-events/ics-advisories/icsa-26-316-05) · [also: CISA](https://www.cisa.gov/news-events/ics-advisories/icsa-26-316-04) · [also: CISA](https://www.cisa.gov/news-events/ics-advisories/icsa-26-316-03)
- [CVE-2026-123158 (\*sshole): OpenSSH Password Authentication Bypass on Alpine Linux](https://www.stepsecurity.io/blog/cve-2026-123158-sshhole-openssh-musl-password-bypass)

### Supply chain and releases (2)

- [PyPI's redesigned project pages with clearer security signals are now live](https://blog.pypi.org/posts/2026-11-06-ui-updates-complete/)
- [Binance’s $6.34 Billion Hack: Developer Machines, AI Tooling, and Wallet Signing](https://www.stepsecurity.io/blog/binance-hack-developer-machine-security)

### Amazon Web Services security bulletins (1)

- [CVE-2026-126793 - Privilege escalation in Aurora PostgreSQL dialect detection in AWS Advanced Go Wrapper](https://aws.amazon.com/security/security-bulletins/rss/2026-182-aws/) · [also: AWS security bulletin](https://aws.amazon.com/security/security-bulletins/rss/2026-181-aws/)

### Malware advisories (1)

- [500 malicious packages added to the GitHub Advisory Database this week](https://github.com/advisories?query=type%3Amalware)

### Quick hits (3)

- [Project Zero will disclose vulnerabilities 90 days after reporting](https://projectzero.google/2026/11/disclosure-policy.html)
- ["Humble Bundle just killed most of their security through vibecoding"](https://forums.spacebattles.com/threads/humble-bundle-just-completely-killed-most-of-their-security-through-vibecoding.1346295/)
- [Turning a 22-year-old Postgres type-confusion primitive into RCE](https://mehmetince.net/part-3-6-how-i-ended-up-turning-a-22-year-old-postgresql-type-confusion-primitive-into-rce-cve-2026-6460/) · [also: filemanager-cdn.mziq.com](https://filemanager-cdn.mziq.com/published/3972906b-e50b-4f74-ab74-4d0d32125d11/9567a559-d20a-486e-a7ae-5c6bbc030c5e)

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.

\[Ada\] AI and agent security. This is the security news for November tenth through November sixteenth, and it opens with Cursor explaining how a trojanized Claude Code extension ended up in its marketplace.

According to BleepingComputer, Cursor published an incident report on Friday confirming a breach that began at Fireworks AI, a third-party provider Cursor used for model training and inference. Cursor says attackers obtained an access key stored at Fireworks and used it to copy data from a storage bucket Cursor used to share information with that provider. The bucket held historical AI coding-agent traces, mostly from twenty twenty-five, from users who had not enabled Privacy Mode. Some of those traces contained unredacted A P I credentials, and the attackers used them to reach other Cursor systems. Cursor says fewer than zero point zero five percent of users had traces exposed, and fewer than zero point zero one percent of customers had A P I credentials compromised. From its software-as-a-service tools, the attackers took archived non-production source code repositories, limited support conversations, product usage metrics and contact details, summaries of some coding-agent conversations, and credentials customers had shared with Cursor's support team.

One of the stolen credentials allowed publishing to Cursor's extension marketplace. On October twenty-ninth, the attackers placed tampered copies of Anthropic's Claude Code extension, versions two point one point three one two and two point one point three one three, on the cache that serves the marketplace. They were available for about eleven hours, and Cursor estimates up to five thousand users may have received them. StepSecurity's analysis says the malicious build is identical to the official extension except for a small loader inside its extension dot J S bundle, which runs as soon as the module loads. It appends an attacker key to the S S H authorized keys file, tries to enable S S H access, and sets up persistence: a cron entry on Linux, a LaunchAgent on macOS, and a scheduled task or Startup folder entry on Windows. BleepingComputer notes that coding errors likely stopped the backdoor on many machines where S S H was not already enabled. Anthropic says its own systems and official channels were not compromised; this affected Cursor users only.

Cursor says it has moved all inference traffic off Fireworks, rotated credentials, rebuilt the marketplace and redeployed production. If you installed or updated the Claude Code extension in Cursor on macOS or Windows on October twenty-ninth, check your authorized keys file and startup items, and follow Cursor's remediation guidance.

\[Dennis\] The same Fireworks AI incident reached Notion. In a notice on its help site, Notion says Fireworks told it of a security incident affecting fewer than one percent of Notion workspaces. Affected customers can download the exposed data from a Security incident review page in their workspace sidebar, and Notion advises rotating any passwords, A P I keys or tokens found in it. Notion also says the data included unattributable Notion AI chat transcripts used for its Insights feature, with no labels for who wrote them or which workspace they came from, and it recommends treating that data as exposed: anyone who pasted credentials into Notion AI during the period should rotate them.

On Hacker News, user underscore f w took issue with the word unattributable, arguing that the contents of a conversation often identify the person or organization even without metadata. User leumon pointed out that OpenRouter labels Fireworks as a zero data retention provider, and asked how retained data could leak at all; user edot added that the main Fireworks bulletin says some customer model weights and training datasets were also exposed.

\[Grace\] StepSecurity has published a full analysis of the compromised LiteLLM release. LiteLLM, maintained by BerriAI, is a Python S D K and proxy that puts more than a hundred model providers behind one OpenAI-compatible A P I, and it often holds provider keys on production gateways. According to StepSecurity, version one point one-oh-seven point oh, published on October twenty-first, shipped malware in all seven compiled wheels and in the official Docker images. The wheels drop three extra files into site-packages, including a dot P T H file and sitecustomize dot P Y, so the payload runs on every Python interpreter start in that environment, with no import of LiteLLM needed. One payload is a credential stealer; the other is a remote code loader that checks in hourly. Both send data to the domain litellm dot dev, which StepSecurity says was registered on October thirteenth.

The root cause is a leaked secret. StepSecurity says a file committed to a public BerriAI-related repository held the master key for an internal agent platform. The commit was overwritten with a force push, but the key was never revoked and kept working for almost five months. With it, the attacker pulled a GitHub organization owner's C L I token, sat inside BerriAI's GitHub organization for eleven days, and built the release from a lookalike repository using BerriAI's own release workflow. The build provenance on the Docker images names that hyphenated lookalike repository as the source, which is how the switch shows up after the fact. The release was downloaded about forty-nine thousand times from Py P I before removal, and the latest and main-stable image tags pointed to the infected image for part of that day.

StepSecurity's advice: if version one point one-oh-seven point oh was installed from a wheel, or you pulled a LiteLLM container on October twenty-first, treat the host as fully compromised, remove the extra files, pin a clean release, block litellm dot dev, and rotate every credential the host could reach, including provider keys stored in the proxy database.

\[Linus\] The F B I has issued a FLASH advisory on ARTEX, the open-source autonomous penetration testing framework that has come up on this show before. According to BleepingComputer, the bureau says multiple threat actors have used ARTEX since at least August twenty twenty-six against financial, government, technology, defense and healthcare organizations, and that some intrusions went from initial targeting to privileged access in under three hours. ARTEX is written in Go and turns an operator's goals into commands for specialized AI agents that handle reconnaissance, exploitation, privilege escalation and data theft. The F B I says it is especially effective against overlooked systems: partner portals, employee applications, legacy services and third-party integrations.

The advisory ties ARTEX to the October ConnectWise Cloud Backup supply-chain attack, in which, per Microsoft's investigation, the China-state-aligned actor Microsoft tracks as Storm zero nine four zero abused application identities and attempted authentication against more than two thousand two hundred fifty downstream Microsoft three sixty-five tenants; Huntress reported at least one hundred sixty-two compromised tenants. It also cites attacks on South Korean financial institutions that exposed personal information of about sixty-seven thousand people. The original developer removed the GitHub repository on October eighth, but forks, archives and a commercial edition remain, and researchers had found more than one thousand six hundred exposed ARTEX consoles by early November.

The F B I recommends securing internet-facing applications, restricting access to cloud administration interfaces, reviewing application permissions and authentication logs, and rate limiting sensitive A P Is, while watching for bursts of automated requests and unexpected privilege changes. The researcher MalwareMustDie also released an open-source ARTEX detection signature pack. StepSecurity's own write-up on ARTEX makes the point that leaked secrets are its fastest way in, describing one agent that went from a GitHub search for leaked secrets to stolen cloud data in five minutes.

\[Ken\] VulnCheck has published its latest tally of vulnerabilities attributed to Anthropic, seven months into Project Glasswing, and argues the results fall short of the launch claims. VulnCheck says it has tracked four hundred forty-two C V Es attributable to Anthropic, about half a percent of the ninety-three thousand four hundred fifty published this year, and only two of them have been exploited in the wild. It says Anthropic's own disclosure ledger lists thirty-seven thousand nine hundred sixty-six findings, of which seven thousand three hundred forty-nine, or nineteen point four percent, are public and just two hundred seventeen are marked fixed. Since September, by VulnCheck's count, nearly twelve thousand findings were added and only fifteen more fixes recorded, while one thousand six hundred fifty-four findings are marked duplicates and four hundred nineteen withdrawn. VulnCheck also cites curl, which it says rejected nineteen of thirty-four published reports from Anthropic's scanner, confirmed six as low severity and left nine awaiting assessment.

The Hacker News discussion pushed back hard. User rurcliped argued that the ledger only covers Anthropic's own open-source disclosure program, while most Glasswing discoveries happen when partners point the model at their own proprietary code. User enthuseyourcurb, who says they work at a large tech company, said they have spent nearly every working day since the model's launch patching serious issues in private codebases that will never get a C V E. User pizlonator, who maintains Fil-C, said all nine reports they received from Anthropic were real bugs with exceptional test cases. User sigmar noted that a low exploitation rate is the intended outcome of patching early. And user delichon worried about a broader pattern of AI being better at finding vulnerabilities than fixing them.

\[Ada\] Three AWS bulletins this week concern Model Context Protocol tooling. AWS API MCP Server, an open-source server that lets AI assistants manage AWS accounts through A P I calls, normalized operation names incorrectly before checking them against the operator's security policy. Operations whose names contain consecutive capital letters never matched a correctly written deny list or elicit list entry, so they could run even where the policy said to block them or ask first. It is tracked as C V E twenty twenty-six, one two five two seven one, rated eight point five, and AWS says there is no workaround other than upgrading to the latest version.

Separately, Amazon's open-source mcp-gateway-registry has two bulletins. C V E twenty twenty-six, one two five nine two six is a server-side request forgery flaw: a user allowed to register or modify a server could point its upstream address at internal services, and in some cases see the responses. C V E twenty twenty-six, one zero zero three zero seven is missing authorization on the external management A P I, letting a low-privileged user map their own identity group to the registry administrator scope. AWS notes the second was fixed in an earlier release, so a deployment can carry that fix and still be exposed to the first. The advice for both is to update to the latest version.

\[Dennis\] Critical vulnerabilities in widely deployed software. Apache HTTP Server has a critical flaw in its HTTP slash two module, mod underscore http two, tracked as C V E twenty twenty-six, one one nine six one eight. According to Aikido, citing Apache's advisory, a race condition in the module's assign headers function lets an unauthenticated remote attacker trigger heap corruption and run arbitrary code with the privileges of the web server process. Anthropic and Google reported it to Apache. Aikido says there were no reports of exploitation in the wild as of publication. Its C V E record rates it eight point one, high, while Apache classes it as critical; Hacker News user acidburnNSA noted it appears to be the first issue Apache has ranked critical since twenty twenty-one.

Affected servers run releases from two point four point fifty-five onward with HTTP slash two enabled; the fix is in two point four point seventy-one. Aikido says not every operating system has a fixed package yet, so check your distribution's tracker. If you can't patch yet, Aikido suggests disabling HTTP slash two by removing the h two and h two c entries and limiting Protocols to HTTP slash one point one. It also says servers behind a standard HTTP load balancer are not exposed in the usual setup.

Two practical notes from Hacker News. User neurostimulant pointed out that many Apache and P H P installations run the prefork multi-processing module and are likely not affected. And user rurcliped warned that if you shut Apache off temporarily, a certbot cron job or systemd timer can restart it, so stopping and disabling the service alone may not keep it down.

\[Grace\] The React team published C V E twenty twenty-six, one two six one eight seven on November tenth, an unauthenticated remote code execution flaw in React Server Components rated nine point five. It affects the three bundler packages that implement the Flight transport: react-server-dom-webpack, react-server-dom-parcel and react-server-dom-turbopack. Fixes shipped the same day across four release lines. StepSecurity's analysis reconstructs the bug: when the Flight writer sends a large typed array, it passes a view over the application's shared buffer instead of a copy, and if a downstream consumer detaches that buffer, the writer silently drops the bytes while the header for that chunk stays on the wire. The parser then reads whatever follows as the promised data, which can smuggle in a module reference and run a module from the application's own bundle.

The GitHub advisory says this only affects extremely rare cases where responses carry multiple typed arrays and an attacker-controlled value, but recommends updating immediately. Client-only React apps without a server are not affected. StepSecurity notes it is the second serious Flight bug in twelve months. If you run React Server Components through Next.js or another framework, update to the patched release.

\[Linus\] Microsoft's November Patch Tuesday was another very large one. According to Krebs on Security, Microsoft fixed one thousand ninety-three vulnerabilities, sixty-three of them critical, and has now shipped more than five thousand one hundred security patches this year, compared with around one hundred a month a year ago; Microsoft credits advances in AI for finding and fixing more bugs. One flaw is already exploited: C V E twenty twenty-six, one two four one nine four, a spoofing bug in Outlook affecting Microsoft three sixty-five Apps for Enterprise and Office. C I S A added it to its Known Exploited Vulnerabilities catalog with a due date of November twenty-fourth.

Google released a Chrome update fixing nine issues, including C V E twenty twenty-six, one two five three five zero, a use-after-free in the X M L component that Krebs describes as Chrome's sixth zero-day this year; it is also in the K E V catalog and affects other Chromium-based browsers. Apple shipped out-of-cycle updates fixing at least nine vulnerabilities, including iMessage flaws that could let an attacker join a group chat or see messages sent before they joined. Lee Neely of the SANS Institute noted Apple backported the fix to iOS eighteen and advised checking every Apple device rather than assuming updates went out automatically. Firefox one sixty fixes almost one hundred C V Es, and Adobe patched at least one hundred ten vulnerabilities with no known exploitation.

For perspective, Chris Goettl of Ivanti told Krebs that of some ninety-five thousand C V Es assigned this year, only two hundred seventeen have been added to the K E V list, and that forty-four percent of this year's exploited vulnerabilities sit in edge devices and I T management tools.

\[Ken\] The SANS Internet Storm Center is seeing exploit attempts against GPUStack, open-source software that manages clusters of G P Us for AI workloads. The flaw, C V E twenty twenty-six, one zero two eight nine five, rated nine point eight and now in C I S A's K E V catalog, is a trivial authentication bypass. According to Johannes Ullrich of SANS, GPUStack trusts a custom header, X dash Gpustack dash Real dash I P, to identify where a request came from, and skips authentication when that header claims the request came from the loopback address. Anyone can send that header. SANS says the scanning source has also been probing for other AI endpoints, such as model listing and messages A P I paths. Older versions are affected; the fixed release came out at the end of June, though the C V E was only assigned on November fourth. If you run GPUStack, update, and keep its interface off the internet.

\[Ada\] C I S A added two chained Barracuda Networks flaws to its Known Exploited Vulnerabilities catalog on November tenth, with a due date of November thirteenth. C V E twenty twenty-six, one one six six zero six, rated nine point three, is an authentication bypass in the Email Security Gateway that lets an unauthenticated remote attacker invoke a configuration import and restore a crafted backup archive. C V E twenty twenty-six, one one one five six zero, rated eight point six, is an operating system command injection that lets an administrator inject commands that run as root through a crafted timezone value; it affects the Email Security Gateway, Load Balancer, Application Delivery Controller, Message Archiver, Web Application Firewall and Web Security Gateway, among others. Chained together, they take an attacker from no credentials to root. If you run these appliances, apply Barracuda's updates.

\[Dennis\] Last month's big flaws are now working their way into industrial products. C I S A republished Siemens advisories this week covering the Brotli decoder flaw known as Node2Shell, C V E twenty twenty-six, one one one four nine seven, and the P H P runtime flaws known as PHP2Shell, C V E twenty twenty-six, one zero nine six one eight and one zero nine six one nine. All three are in the K E V catalog. Siemens lists the Desigo C C building management family and Gridscale X grid software as affected by Node2Shell, and says it has released new versions. The RUGGEDCOM A P E eighteen oh eight, which runs Palo Alto Networks PAN-OS, is affected by both; Siemens points customers to Palo Alto's own fixes and workarounds. Industrial Edge Management is under assessment, with mitigations for setups that have management applications installed. C I S A's usual advice applies: keep control system devices off the internet and behind firewalls.

\[Grace\] A follow-up on last week's OpenSSH flaw on musl-based systems, nicknamed sshole and tracked as C V E twenty twenty-six, one two three one five eight. C I S A has added it to the K E V catalog, and StepSecurity reproduced it in a lab on an Alpine virtual machine and an Alpine container with port twenty-two published. StepSecurity says the exposure is narrower than the headline: a host is exposed only if it runs an unpatched sshd on musl without PAM, has password authentication enabled, has an account allowed to log in whose shadow password field is an asterisk, and has at least one real SHA-crypt hash on the system. It also warns that version-based scanners produced false positives after Alpine's backported fix. Public proofs of concept exist, so expect scanning. Update OpenSSH, then restart sshd, especially in Alpine images built with sshd for remote debugging.

\[Linus\] Supply chain and releases. Py P I's redesigned project pages are now live, with a new security tab built around release provenance. According to the Py P I blog, each release now shows whether its files carry attestations: Py P I Publish attestations, which confirm a file came through a Trusted Publisher, and SLSA build provenance, which links a file to its source commit and build platform. A release with no attestations shows as having no provenance, which the Py P I team stresses does not mean it is malicious, only that its origin cannot be cryptographically verified. More useful are the new warnings: incomplete provenance, when some files are attested and others are not, and inconsistent provenance, when files in one release come from different build systems or repositories. Those are the signals that would show up if an attacker took over an account or a publishing workflow.

On Hacker News, user cowlevel noted this helps detect hijacked releases of benign projects, not packages that were malicious from the start. User pasteleft asked whether pip shows these signals at install time, since most people never open the project page. User eesmith asked how independent platforms such as Codeberg or Sourcehut could become trusted publishers, and user mhils pointed to the open request for trusted publishing to support reusable workflows.

\[Ken\] StepSecurity has summarized Binance's November third initial findings on its six point three four billion dollar theft. According to that summary, the attack began when an engineer ran a malicious coding assignment, presented as an interview task, on a work computer. A malicious pyenv component ran commands and contacted an external server; the attacker then used an internal AI tool to reach other engineers' computers and obtained credentials with container deployment permissions, which led to wallet signing. Binance has not named the AI tool or explained how it was used. The F B I has attributed the theft to North Korea's TraderTraitor activity. Binance reports about two point eight five billion dollars frozen or intercepted, and says it has rebuilt the affected wallet system and the internal AI tool. StepSecurity's advice is to review what a compromised developer machine could reach: local credentials, shared tools, deployment permissions and signing services.

\[Ada\] Amazon Web Services security bulletins. Two new bulletins this week. C V E twenty twenty-six, one two six seven nine three affects the AWS Advanced Go Wrapper: an untrusted search path issue in Aurora PostgreSQL dialect detection could let an authenticated database user plant a function that runs when another user connects through an affected wrapper, escalating to that user's privileges, including members of the R D S superuser role. C V E twenty twenty-six, one two seven one one nine affects the E M R module of the AWS S D K for pandas, also known as awswrangler: when a cluster is created without an explicit log location, the library uses a predictable default S3 bucket name without checking that the caller's account owns it, so someone who creates that bucket first could read the cluster logs, including job step output. For both, update to the fixed release, and set an explicit log bucket for E M R clusters.

\[Dennis\] Malware advisories. The GitHub Advisory Database added five hundred malware advisories this week: four hundred eighty-nine for N P M packages and eleven for Py P I. Much of the N P M volume follows one pattern, packages with long names styled as ebook downloads, each flagged as fully compromising any machine that installs it. On the Py P I side, the advisories describe a cluster of packages named after the Gate dot io exchange, such as gateio-client and gatepay-sdk-python, that exfiltrate basic host information on install or import. Axiom-dragon-xyz steals cryptocurrency wallet files and cloud credentials during installation, and ps-bool exfiltrates photos and documents to a Telegram bot from Android devices. GitHub's advice for the worst of these is to treat the machine as compromised and rotate secrets from a different computer.

\[Grace\] Quick hits. Google Project Zero has changed its disclosure policy. Since twenty twenty-one it has used ninety days plus thirty days of patch soak time. Starting now, the policy is Full ninety: details go public ninety days after a report, whether or not the bug is fixed. It moves to sixty days on February second, twenty twenty-seven, and to thirty days on January third, twenty twenty-eight. Project Zero says language models have sped up both discovery and patch reverse engineering, to the point where zero time from patch to exploit is a reasonable model, and that it now finds several vulnerabilities a week in targets where it once found a few a year. Grace periods and most special cases are gone; exceptions remain for bugs exploited in the wild, already public bugs, and vendor requests. On Hacker News, user Aissen wondered how this fits Android's long delay between patching and public disclosure, and user PeterStuer argued that on-premises systems without automatic updates will struggle to keep up.

\[Linus\] A thread on the SpaceBattles forums claims that Humble Bundle's rebuilt website weakened account security, and it drew wide discussion on Hacker News. User dtech quoted the thread's claim that a password can now be changed without entering the old one or confirming the new one. User Ekaros and user trvz described the new email code at every login, which Ekaros said arrived an hour late; user oliwarner said logging in failed through both Google and email. The claims come from users, not from Humble Bundle.

\[Ken\] Researcher Mehmet Ince has published the third part of a series describing how he turned a type-confusion primitive in PostgreSQL core, which he describes as twenty-two years old, into remote code execution, triggerable by the lowest-privileged login role. He says the fixes landed on November ninth in a PostgreSQL release that addressed fifty-two C V Es; if you run PostgreSQL, update to the latest minor version. And in Brazil, according to a disclosure document the company published, the insurer Caixa reported a breach that exposed the names and tax I Ds of all seven point six million of its customers.

\[Linus\] That concludes RobotPaper's weekly AI security news. The theme this week was trust in the tools developers rely on: a provider breach that turned into a trojanized extension, a package release built from a lookalike repository, and AI agents that turn leaked secrets into access within minutes. The habits stay the same: revoke leaked keys at the source, keep agent traces and credentials apart, check provenance before you install, and patch internet-facing servers quickly. 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.
