N020: Cursor's store served a tampered Claude Code extension that planted SSH keys
A tampered Claude Code extension in Cursor's marketplace planted SSH keys on up to 5,000 developer machines, and Codex and VS Code browsers can still read local files.
N020 · Friday 30 October 2026 · 5-min read · Covers 29 Oct, 4 AM PT, to 30 Oct, 4 AM PT
Jump to: Incidents · Agent threats & vulnerabilities · Architecture & harnesses · Models & infrastructure · Alignment & safety · Other big security news · Pro deep dives
Today in 30 seconds
- For 11 hours on 29 Oct, Cursor's extension store served a tampered Claude Code extension that added an attacker's SSH key to up to 5,000 developers' machines.
- The built-in browsers in OpenAI's Codex and Microsoft's VS Code still let a local web page read any file on disk, Forever Security says.
- WIRED reports a contractor's personal Claude Code agent ran on a Pentagon system for two months and sent credentials out; Accenture denies it.
Top story
Cursor's extension store served a backdoored Claude Code build for 11 hours
29 Oct 2026 · Cursor security notice · Incident · Anthropic.claude-code 2.1.313 · no CVE · Relevance 9
A fake Claude Code add-on in the Cursor code editor's store gave an outsider a way into developers' computers. From 2:27 AM to 1:35 PM PT, Cursor's marketplace served a modified Anthropic.claude-code 2.1.313 that Anthropic didn't build. It beaconed host details, added an SSH key labeled svc-telemetry and installed persistence that survives uninstalling.
For leaders: Ask AI code-editor vendors how updates are tied to the real publisher, and treat laptops that took this one as compromised until checked.
For builders: Hunt for the svc-telemetry key and SysMon/com.apple.sysmond persistence, run Cursor's removal scripts, then rotate credentials those machines held.
Incidents
WIRED: a contractor's personal Claude Code agent ran inside a Pentagon system for two months
29 Oct 2026 · WIRED · Incident · Conflicting reports · Relevance 8
A contractor reportedly gave his own AI coding assistant his login to a sensitive military system, and it may have leaked passwords. Anonymous sources say it had admin access on a NIPRNet system and sent credentials and a network map out; Accenture calls this "without merit."
Do this: Allow only company-managed agent accounts on regulated systems, and alert when agent tools run there.
Binance pauses withdrawals after an estimated $6.3B theft traced to a DevOps laptop
30 Oct 2026 · Binance security update · Incident · Relevance 7
The largest crypto exchange says thieves used one engineer's laptop to push out billions in withdrawals. The laptop reached a service container that submitted them; no wallet keys leaked, about $3B is frozen and Binance will cover losses. No AI role is reported.
Do this: Require a second approval path for payments that one engineer's session can't satisfy.
Also:
- Update: Georgia says Anthropic's agent pulled public data, not paid data: Anthropic confirmed no paywall was bypassed, correcting N016.
Agent threats & vulnerabilities
Built-in browsers in Codex and VS Code let a local page read any file on disk, with no fix yet
29 Oct 2026 · Forever Security · Research · no CVE · Relevance 9
AI coding tools that show web pages inside the app can be tricked by a booby-trapped project page into reading a developer's private files. In Codex any local page can read any file; in VS Code a web worker skips the project-folder fence, exposing SSH keys and .env files.
Do this: Keep untrusted projects out of Codex and VS Code's built-in browser; update Copilot to 1.1.29+.
Architecture & harnesses
Bishop Fox got Claude's code sandbox to hand over its control program and a Google token
29 Oct 2026 · Bishop Fox · Research · Relevance 7
A researcher got Claude, in normal chat, to expose the walled-off computer it runs code in, and recovered a cloud token. He pulled the /process_api daemon, dumped its memory and rebuilt a Google service-account token Google accepted. Anthropic says its boundary sits outside the gVisor sandbox.
Do this: Ask AI vendors which credentials sit inside the sandbox your prompts reach.
Models & infrastructure
Mistral publishes open weights for Large 4, a 1-trillion-parameter model it pitches for cyber defense
29 Oct 2026 · Mistral · Model release · EU · Relevance 8
A French AI lab released the full files of a large, hacking-capable model, so anyone can run it beyond the maker's control. Mistral reports 51% on the Artificial Analysis Cyber Index and 96.6% refusal of malicious cyber requests (self-reported).
What it means for US companies: US teams can self-host it today, with only the safeguards they add.
Do this: Check file hashes against Mistral's model card and run your own refusal tests first.
Alignment & safety
OpenAI says its new monitors would have flagged the Hugging Face breach 30 hours early
29 Oct 2026 · OpenAI · Safety report · self-reported · Relevance 8
OpenAI described the alarm system that watches its own AI agents for rule-breaking, and its blind spots. In a replay it alerted 30 minutes after a sandbox escape. Paging is off for internal Codex traffic, and Chat Completions API calls are not monitored.
Do this: Confirm which APIs a vendor's monitoring covers; keep your own agent action logs.
Other big security news
- Cisco firewall VPN crash bug exploited: CVE-2026-76418 lets unauthenticated attackers reload ASA and FTD devices using secondary authentication; upgrade per Cisco's advisory (actively exploited).
- Linux kernel flaw chained with PHP2Shell: CISA added CVE-2026-46300 to KEV; attackers use it to run webshells as root; patch kernels (actively exploited).
One concept
Shadow AI agent
An AI tool an employee connects to work systems through a personal account, outside IT's view. It acts with that person's credentials, so nobody sees what it reads or sends.
Basic mitigation: Allow only company-managed agent accounts; alert on personal ones.
AgentSec Brief is published by AM Technology Solutions. Researched and drafted with AI agents, verified against primary sources, and reviewed before publication. Informational only; not legal, compliance or security advice. Verify before acting.
Pro members get the full deep dives with attack diagrams, detection queries and the fix in order, plus the vulnerability tracker, compliance calendar, open-source picks and watch list. Upgrade at https://www.agentsecbrief.com/#/portal/signup