How to Audit Your IDE Extensions After the Cursor Claude Code Backdoor Incident
๐
November 2026๐ AI Dev Toolsโฑ 7 min read
How to Audit Your IDE Extensions After the Cursor Claude Code Backdoor Incident
Quick answer: On October 29, 2026, a malicious version (2.1.313) of Anthropic's official Claude Code VS Code extension was distributed through Anthropic's own verified account in Cursor's extension marketplace, affecting at most 5,000 users before Cursor removed it and published a security notice the same day (updated October 30 with expanded indicators). The payload tried to plant an SSH backdoor with persistence and beaconing. If your team uses Cursor or VS Code-based AI editors: confirm you were never on 2.1.313, check
~/.ssh/authorized_keys and your LaunchAgents for intruders, don't assume "official publisher" means "safe" anymore, and put an extension allowlist plus a periodic audit on the calendar. This guide gives you the exact checks.
---
What actually happened
On Thursday,
October 29, 2026, the malicious version
2.1.313 was published to
Anthropic's official, verified publisher account on the Cursor marketplace โ not a lookalike. It effectively superseded the prior 2.1.312, and because Cursor auto-updates extensions by default, many users received it without clicking anything.
Cursor disclosed the incident on October 30; the details, corroborated by
Ars Technica, security firm
Phorion, and researcher Kevin Beaumont, matter for every development team:
- Cursor estimates at most 5,000 users were exposed. Anthropic pushed a clean 2.1.314 several hours later; in its October 30 follow-up, Cursor also flagged 2.1.312 as compromised, so treat either version as suspect. As of this writing, neither company has publicly explained how the verified account was hijacked.
- The backdoor was cross-platform (Windows, macOS, Linux). Its capabilities, per Cursor's disclosure: collect host and user reconnaissance (hostname, username, OS, architecture, home directory, environment variable names, working directories, Git email, public IP); attempt to append an attacker-controlled SSH key to
authorized_keys, enable the local SSH service, and open a reverse SSH tunnel; make the tunnel persistent via a macOS LaunchAgent or a Windows scheduled task with a Startup-script fallback, retrying roughly every 60 seconds; and report status via HTTP POST to /beacon on port 8099.
- The malware was buggy. Coding errors in the Windows and macOS paths meant the SSH-key and tunnel steps often failed unless Cursor had been started with elevated privileges. Phorion's assessment: despite the flaws, it still installed persistence and beaconed reconnaissance โ including environment variable names, which is enough to inventory which machines hold cloud credentials and API tokens worth attacking next.
Exposure window: anyone who installed or updated the extension from Cursor's marketplace between roughly
5:27 a.m. and 4:35 p.m. ET on October 29 (2:27 a.m.โ1:35 p.m. PT per Cursor's own notice) should assume they ran the malicious build. Cursor worked with the cloud provider hosting the command server to sinkhole it, removed the bad version, then shut the marketplace entirely for a short period to rotate keys and audit access controls before reopening.
One more uncomfortable fact: Ars reported that Anthropic had not issued a public customer warning about the incident as of the disclosure date. Do not count on vendors to notify you promptly โ the checks below are yours to run.
Why "it was from the official account" is the scary part
Most developer security habits assume the marketplace layer is trustworthy: verify the publisher name, check install counts, done. This attack invalidated that assumption. The attacker gained access to the official distribution channel itself, so every signal users rely on โ verified badge, official publisher, plausible version number โ pointed the wrong way.
The structural lesson mirrors the Postman/Claude Code plugin affair from the same week:
AI coding tools concentrate extraordinary privilege (your source code, your credentials, your production paths), and the supply chain feeding them โ extensions, plugins, MCP servers, hooks โ is now a target. Teams need a repeatable audit routine: know where the doors are, check them on a schedule, and keep the list of extensions you trust as short as possible.
Immediate triage: were you exposed?
1. Find out whether you ever ran 2.1.313. In VS Code, open the Extensions view, click the Claude Code extension, and check the version history / "Switch to..." version list. In Cursor, check your extensions folder directly:
- macOS/Linux:
~/.cursor/extensions/ โ look for the anthropic extension directory and inspect its package.json for a version field. Any machine that shows 2.1.312 or 2.1.313, or that updated the extension on October 29 within the exposure window, should be treated as potentially compromised even after updating to 2.1.314 โ updating to the clean version does not remove an existing infection.
2. Check for the persistence and access artifacts. On every machine in scope:
~/.ssh/authorized_keys: every entry should trace to a key you generated or were explicitly given. An unfamiliar key appended around October 29โ30 is the incident's signature move.
- macOS:
~/Library/LaunchAgents/ for a plist you don't recognize that references Cursor, node, or a script retrying a network connection.
- All platforms: look for outbound connections to port 8099, and any local process listening or beaconing unexpectedly. Cursor and Phorion both published indicators of compromise (IPs and hashes) โ compare against those lists rather than guessing.
- Review shell history and recent logins (
last on macOS/Linux) around the exposure window for anything unexplained.
3. If you find artifacts: rotate, don't just clean. Assume cloud credentials and API tokens accessible on that machine are burned. Rotate AWS/GCP/Azure keys, GitHub tokens, npm tokens, and any SSH keys that lived there. Rebuild the machine if you can't establish a clean baseline โ a persistence implant that survived one cleanup attempt is cheaper to wipe than to chase.
Ongoing controls that would have blunted this attack
Turn off silent auto-update for extensions, or at least gate it. Cursor (and VS Code) updating extensions automatically is precisely how 2.1.313 reached machines with nobody clicking anything. Auto-update from a verified publisher is convenient; the trade-off is that a single compromised account gets pushed to your entire team within hours. Options: disable auto-update and run updates on a schedule you control, or enable it but pair it with the audit routine below.
Run a recurring extension audit โ monthly is enough. Ten minutes per machine:
- List every installed extension (
code --list-extensions / the equivalent for your editor). Each entry should have an owner on your team who can say why it's there.
- Verify publishers for anything with file-system, terminal, or network reach. Note that publisher verification failed in this incident โ so also check when the extension last updated and whether that update aligns with your own install/update actions.
- Delete extensions nobody owns. The biggest risk reduction is a shorter list.
Prefer allowlisting over blocklisting. For teams, an internal allowlist (even an informal tracked list in a repo) beats reacting after each incident. New extension requests go through a quick review: what does it do, what permissions does it need, who maintains it, what happens if its publisher account is compromised. That last question sounds paranoid; it is now the literal attack path.
Treat your IDE like the privileged environment it is. Extensions run with your user's permissions โ they can read your source, your environment, and your credential stores. Apply the same discipline you (hopefully) apply to CI runners: minimal extensions, separate credentials per machine where feasible, and no long-lived production tokens on dev laptops.
Watch the whole supply chain, not just extensions. The same week, the Postman desktop app was caught silently installing plugins into Claude Code and Codex and attempting to approve its own trust hooks. Extensions, agent plugins, MCP servers, and hooks are all channels by which third-party code reaches your AI tooling. If you run AI coding agents, audit those surfaces too โ our companion guide on auditing Claude Code and Codex plugins covers the agent-side checklist.
The honest bottom line
This incident's impact was limited by luck as much as defense: the malware had bugs, the exposure window was about 11 hours, and the command server got sinkholed. Next time those mitigations may not exist. The verified-publisher model is broken for AI dev tools until marketplaces add real signing and provenance guarantees; until then, "official account" is a necessary but nowhere near sufficient signal. Assume any extension can turn hostile on any update, keep your installed set small, audit it monthly, and have a rotation plan for the credentials on your dev machines. Ten minutes of checklist a month is cheap insurance against becoming the next 5,000.