A Backdoored Claude Code Extension Was Live for 11 Hours. Here’s the One-Line Check.
TL;DR: For about 11 hours on 29 October, two tampered builds of the Claude Code extension were available in Cursor’s own extension marketplace. They planted an SSH key on the machine, opened a tunnel back to the attacker and installed startup persistence. Cursor says at most 5,000 users were exposed, and that the bad builds were not made by Anthropic. The part that catches people out: uninstalling the extension does not remove the backdoor. There is a one-line check below.
What happened to the Claude Code extension
On 29 October 2026, Cursor published a security notice saying that malicious copies of the Claude Code extension (`Anthropic.claude-code`) had been served from its own VS Code extension marketplace “between about 2:27am PT and 1:35pm PT” that day — roughly eleven hours, or 09:27 to 20:35 UTC in Sophos’s framing. An update to the notice on 30 October widened the list to two versions, 2.1.312 and 2.1.313, in the macOS and Windows builds.
The teardown came from Phorion, where Calum Hall found that only one file had been altered. A block of code had been pasted in front of the real Claude Code extension, so it ran the moment the editor loaded the file — before the extension was even switched on. No chat, no prompt, no click.
What it then did, per Cursor’s and Phorion’s matching accounts:
- Added its own SSH key — labelled `svc-telemetry` — to the machine’s authorised keys, and tried to switch remote login on.
- Opened a reverse tunnel out to the attacker’s server at `134.209.82.225`, retrying roughly every 60 seconds and carrying on after the editor was closed.
- Installed persistence that survives a reboot: a LaunchAgent disguised as `com.apple.sysmond` on macOS, a scheduled task called `SysMon` on Windows.
- Phoned home with the hostname, username, operating system, home and working directory paths, Git email and public IP — plus the names of up to 40 environment variables matching patterns like `KEY`, `TOKEN`, `SECRET`, `AWS` and `GCP`. Names, not values. A shopping list rather than the shopping.
Cursor pulled the bad Claude Code extension builds, then shut its whole marketplace down “out of an abundance of caution”, audited access, “confirmed no other extensions were tampered with”, rotated its keys and rebuilt. It also worked with the hosting provider to take the attacker’s server offline, so the tunnel and the beacon no longer connect to anything.
Two things worth being precise about. Cursor states plainly: “These malicious versions of the extension were not developed by Anthropic”, and that it has no evidence other versions or the official distribution were affected. An Anthropic engineer said much the same in the Hacker News thread, which ran to 876 points: “This only impacted the extension as it was hosted on the Cursor marketplace.” Microsoft’s Visual Studio Marketplace — the default store in plain VS Code — has had no malicious version reported, per Help Net Security. And nobody has yet explained publicly how a tampered build got in. Early reports pointed at the Open VSX registry and were later corrected; Anthropic has published nothing about the incident on its news page as of 31 October.
Why it matters if you don’t code
Installing an extension is a click. That is the whole point of extensions, and it is why this lands on people who have never read a stack trace.
Three things follow.
You did not have to choose the bad Claude Code extension. Editors update extensions quietly in the background. If your laptop happened to fetch an update inside that eleven-hour window, you got it.
What was at risk was your laptop, not your app. Most AI-coding scares are about the thing you built. This one was about the machine you built it on — the browser sessions, the saved keys, the password manager you had already unlocked. Our piece on why API key security is really about who holds a copy applies directly here.
Removing the Claude Code extension does not undo the damage. Both Cursor and Sophos say so outright. An SSH key sitting in a file, and a scheduled task, do not care that the editor is gone.
How to check whether you had the bad Claude Code extension
Five minutes, in this order.
- Run the one-line check. On macOS, open Terminal and paste Cursor’s command: `grep -qs svc-telemetry ~/.ssh/authorized_keys && echo “svc-telemetry key found” || echo “No key found; machine not impacted”`. On Windows, Cursor’s notice has the PowerShell equivalent. No key found means no impact.
- If it finds something, stop using the machine for anything sensitive. Do not log into your bank or your hosting account from it.
- Run Cursor’s clean-up script — `remediate-macos.sh` or `remediate-windows.ps1`, both linked from the notice. They exit with an error if anything is left behind.
- Rotate your credentials from a different, trusted device — AI provider keys, Stripe, your database, GitHub. Treat anything the laptop could reach as known. A manager that holds keys for agents, like 1Password’s, makes this less painful next time.
- Reinstall the Claude Code extension from the official marketplace, and turn off automatic extension updates if your editor lets you. One Hacker News commenter put the mood well: everyone is going to set their update delay to the maximum after this.
Our vibe coding glossary covers SSH, persistence and the rest in plain English.
Who should care (and who shouldn’t)
- Building a revenue app: care most. Run the check, then rotate every key the machine could touch. Your customers’ data sits behind those keys.
- Internal tool for your business: this is an IT matter now, not a coding one. Tell whoever manages your company’s laptops today — Cursor’s indicator list is written for exactly that person.
- Website or landing page: if you never installed the Claude Code extension in Cursor, there is nothing on your machine to clean. Turn off silent extension updates anyway.
- Learning to build: do not let this put you off. It is an argument for sticking to one editor and one official store, not for giving up.
- No idea yet: start with the 60-second Tool Finder quiz — it will steer you towards tools whose risks you can actually manage.
Our take
The uncomfortable detail is not the backdoor in the Claude Code extension; it is the signing. As one commenter in the thread pointed out, VS Code extensions need not be signed at all — so a marketplace is a trust decision you make once and then never revisit. We would not drop Cursor over this: its response was fast, specific and unusually candid, which is more than most vendors manage. But we would stop treating “it came from the marketplace” as a safety claim, and we would turn off automatic extension updates on any machine that holds real credentials. Convenience was the attack surface here.
Not sure which AI coding tool suits you? Take the 60-second Tool Finder quiz →
Some links on this page are affiliate links. If you buy through them we may earn a commission at no extra cost to you.
FAQ
Which versions of the Claude Code extension were affected?
Cursor’s notice names Claude Code extension versions 2.1.312 and 2.1.313 (`Anthropic.claude-code`), in the macOS and Windows builds, downloaded from Cursor’s own marketplace between roughly 2:27am and 1:35pm PT on 29 October 2026. Cursor says it has no evidence that other versions or the official distribution were affected.
Is uninstalling the Claude Code extension enough?
No. Both Cursor and Sophos say removing the extension leaves the SSH key, the startup persistence and the tunnel process in place. You need to run the check, then Cursor’s clean-up script, then rotate credentials from a trusted device.
Was Anthropic itself hacked?
Cursor states the malicious builds “were not developed by Anthropic” and that the official distribution shows no sign of being affected; an Anthropic engineer said only the Cursor-hosted copy was involved. How the tampered build reached the marketplace has not been explained publicly.
