Build · founder · 6 min read
The Claude Code Extension Was Trojanized on Cursor: Check Your Laptop
A poisoned update to Anthropic's Claude Code extension hit Cursor's marketplace on Oct 29. Here is how to tell if you were exposed and what to do.
On October 29, version 2.1.313 of Anthropic’s official Claude Code extension was served to Cursor users with a backdoor inside it. It was live for roughly 11 hours. If you use Cursor, installed or updated extensions that day, and haven’t checked your machine, this is the post to read before your next coffee.
The short version: it was the real extension under the real publisher name, not a lookalike. That’s what makes it nasty. Nobody had to mistype anything.
What happened
Per The Hacker News and other security outlets, a malicious update to anthropic.claude-code appeared on the Cursor Marketplace. The window ran from about 09:27 to 20:35 UTC on October 29 (2:27 a.m. to 1:35 p.m. Pacific). Cursor says at most 5,000 users were affected and that it will contact them directly.
Update, November 14: Cursor has now explained how the attacker got in. A key it had issued to its vendor Fireworks AI was exposed, which led to credentials that reached the cache serving Cursor’s marketplace. Cursor’s notice also now lists two tampered versions, 2.1.312 and 2.1.313. See our breakdown of Cursor’s report.
Cursor said the malicious build was not developed by Anthropic. On October 30, Cursor temporarily shut its marketplace, pulled the bad version, and Anthropic published a clean 2.1.314.
An early version of the coverage said Open VSX was also compromised. That was corrected: as of the latest reporting, only the Cursor Marketplace was affected, and there’s no evidence the update reached the Microsoft VS Code Marketplace.
What the backdoor did
When the editor started and the extension’s panel opened, a small loader ran before the legitimate code. On Windows, macOS and Linux it:
- Added an attacker’s SSH key, labeled “svc-telemetry,” to your
~/.ssh/authorized_keysfile. - Tried to turn on SSH access and open a reverse tunnel to an outside server, so the attacker could reach in rather than you having to open a door.
- Installed persistence so it survives a restart: a scheduled task, a crontab entry, or a LaunchAgent, depending on your system.
- Sent basic details about your machine to the attacker over plain HTTP.
Those details include your public IP, hostname, username, home folder, Git email, and the names (not the values) of up to 40 environment variables containing words like KEY, TOKEN, SECRET, PASS, AWS or GCP. That list is a shopping list: it tells the attacker which of your accounts are worth going after.
No follow-on malware has been observed so far, but a working tunnel means there didn’t need to be one yet.
What we don’t know
Cursor says that on Windows and macOS the SSH step likely fails because the extension usually lacks admin rights, so the attacker probably couldn’t log in through the tunnel there. Treat that as reassuring, not conclusive. The key, the persistence and the data collection still happen, so the cleanup below applies either way.
How to check if you were exposed
You’re in the risk group if you use Cursor and the Claude Code extension auto-updates, and your editor was open at any point on October 29 in that window. Do these in order:
- Check the extension version. Open Cursor’s extensions panel, find Claude Code, and confirm you’re on 2.1.314 or later. If you see 2.1.312 or 2.1.313, treat the machine as compromised.
- Look at your SSH keys. Open
~/.ssh/authorized_keys(on Windows, check your user’s.sshfolder). Delete any line labeled “svc-telemetry.” You didn’t add it. - Look for persistence. Check scheduled tasks on Windows, your crontab on Linux, and
~/Library/LaunchAgentson a Mac for anything you don’t recognize, especially anything created October 29 or later. - Check network logs if you can. The attacker’s address was 134.209.82.225. If you have a router or firewall log, search for it. Most solo founders don’t, and that’s fine.
If you aren’t comfortable poking around system folders, this is the moment to ask a technical friend for 20 minutes. Just don’t skip it.
What to rotate
Assume anything that lived in your environment variables was seen by name, and possibly more if the tunnel was used. Start with the highest-value items:
- Cloud and hosting keys (AWS, Google Cloud, Vercel, Railway, and so on).
- Payment keys, especially Stripe secret keys.
- Database passwords and Supabase service-role keys.
- API keys for AI providers, since a stolen key is a bill you didn’t authorize.
- GitHub tokens and any deploy keys.
Rotate means create a new one, update your app, then revoke the old. Do it in that order or you’ll take your own site down.
The bigger lesson for non-technical builders
Your editor and its extensions run with your permissions, on the same laptop as your email, your banking tab and your production keys. An extension is not a toy; it’s software with the keys to your house.
A few habits cut the risk without turning you into a security engineer:
- Turn off extension auto-update in your editor, or at least delay it a few days. This attack worked because updates arrive silently within hours.
- Keep secrets out of your shell environment where you can. Use your host’s secret manager instead of exporting keys into every terminal.
- Use scoped, short-lived keys for tools that touch production, so one leak doesn’t mean total access.
- Keep your build machine boring. Fewer extensions, no random installs, and no personal accounts logged in on the browser profile you code in.
This is the second time this year an official-looking editor extension has been the way in; we covered the earlier GitHub case in our May write-up. The pattern is repeating because extension marketplaces are a soft target and AI coding tools made them far more popular. Expect more of these, and plan like you will.
Bottom line
Check your version, check authorized_keys, rotate your secrets, and turn off silent auto-updates. Fifteen minutes now beats a very bad week later.
Related guides
founder · 8 min read
10 Apps Non-Technical Founders Built With Vibe Coding
Real examples of SaaS products, marketplaces, and tools built without code: what they built, which tools they used, and what it cost.
founder · 8 min read
Researchers put a virus in an agent's memory file. It spread to the next agent.
New Anthropic/EPFL research says AI agents can pass self-replicating instructions through files like CLAUDE.md. The fix is one paragraph.
founder · 9 min read
Your Agent Can Now Spend Your Money. Nobody Can Prove You Said Yes.
AWS shipped autonomous agent payments this month. The infrastructure works. The paper trail that would settle a disputed charge doesn't exist yet.
Enjoying this guide?
Get weekly practical guides, plus tool updates and implementation playbooks.