Cursor has published the results of its two-week investigation into the October 29 incident in which trojanized copies of Anthropic's Claude Code extension appeared in its VS Code extension marketplace. In a post dated November 13, the AI code editor's maker says the chain began outside Cursor: an access key it had issued to its inference provider Fireworks was exposed at Fireworks, and an attacker used it to copy data from a read-only storage bucket the two companies shared.
That matters because the bucket held coding-agent traces, mostly from 2025 and only from sessions that were not in Cursor's Privacy Mode, and Cursor says some of those traces contained unredacted credentials. The attacker used them to get into SaaS tools Cursor uses, and from there reached more internal data and, eventually, the marketplace cache. It is a clear example of how logs from AI coding agents can become a secrets store by accident.
The chain, as Cursor describes it
| Step | What happened (per Cursor) |
|---|---|
| Before Oct 29 | An access key Cursor issued to Fireworks is exposed by Fireworks; the attacker copies agent traces from a shared bucket |
| — | Credentials found in those traces give access to SaaS tools used by Cursor |
| Oct 29 | Using a stolen credential, the attacker places malicious copies of the Claude Code extension in Cursor's marketplace cache; they stay up about 11 hours |
| Oct 30 | Cursor publishes a security notice with indicators of compromise and check scripts, and starts notifying users |
| Nov 1 | Last attacker activity Cursor has seen; it was blocked |
| Nov 13 | Cursor publishes its post-mortem |
Cursor's post does not say when or how Fireworks exposed the key. AI Weekly and a DEV Community write-up of the post both note that neither carries a statement from Fireworks. Fireworks has separately disclosed its own October intrusions, which we covered here; Cursor's account does not tie the key exposure to a specific one of those incidents, and we don't either.
What was taken
Besides the agent traces, Cursor lists:
- archived, non-production source-code repositories, some with business contact details;
- limited customer-support data and email conversations, including some credentials users had sent to support;
- product usage metrics and contact details;
- summaries of some coding-agent conversations.
Cursor's numbers: at most 5,000 users who requested the specific cached versions on October 29 may have received the malicious build; fewer than 0.05% of users had agent traces exposed; and fewer than 0.01% of customers had API credentials exposed. Affected users were notified directly, the company says.
Cursor says Privacy Mode prompts and code, live customer data and traffic, production systems, the Cursor app and its official updates, release-signing keys, customer code repositories and Cursor Marketplace plugins were not affected.
What the fake extension did
Cursor's new post doesn't describe the payload, but its earlier notice and outside analysis did. According to SecurityWeek's November 2 report, versions 2.1.312 and 2.1.313 of anthropic.claude-code ran JavaScript on load that collected the hostname, username, OS, architecture, working directory and global Git email, then tried to install an SSH key, enable SSH, open a reverse SSH tunnel and set up persistence on macOS and Windows. Cursor said enabling SSH needs admin rights the extension lacked on most machines, so that step usually failed.
Anthropic said it did not build or publish those versions, and that its official channels were not affected; SecurityWeek reports they never reached Microsoft's Visual Studio Marketplace, Open VSX or npm, and Anthropic pushed its legitimate extension straight to 2.1.314 to avoid confusion.
What changed at Cursor
Cursor says it moved all inference off Fireworks and deactivated the key, rotated every exposed credential it found plus all internal credentials, validated every other marketplace extension and rebuilt the store, redeployed production infrastructure, added secret scanning across code repositories and other internal data, tightened network restrictions with providers and brought in outside forensics. Some features were briefly disrupted while services were disabled.
What to do
- Check your extension history. If you installed or updated to Claude Code 2.1.312 or 2.1.313 from Cursor's marketplace around October 29, follow Cursor's October security notice and run its check script. Cursor has warned that uninstalling alone may not be enough because the payload drops standalone persistence scripts.
- Look for the backdoor's leftovers: an unfamiliar key in
authorized_keys, unexpected SSH services, reverse-tunnel processes and new startup items. Cursor's earlier guidance was to isolate affected hosts and consider rebuilding them. - If Cursor emailed you about exposed credentials, rotate exactly what the notice lists — and anything else you ever pasted into an agent session or a support ticket.
- Treat agent logs as sensitive. If you use any AI coding tool without a privacy or zero-retention mode, assume secrets that appear in a session can end up in vendor storage. Keep keys out of prompts and use a secret manager or environment injection instead.
- Questions go to security@cursor.com, per the post.
The open question is Fireworks' side: how Cursor's key was exposed, and whether other customers' keys or shared buckets were reachable the same way.
Sources
- Cursor — An update on our recent security incident cursor.com
- SecurityWeek — Malicious Claude Code Extension Served in Cursor Supply Chain Attack securityweek.com
- AI Weekly — Cursor confirms Fireworks leak enabled malicious Claude Code aiweekly.co
- DEV Community — Cursor's Security Incident, Explained: A Leaked Vendor Key, a Poisoned Extension, and 5,000 Users at Risk dev.to