Cursor Security Incident Update: Fireworks Key Led to Tampered Claude Code Extension for Up to 5,000 Users
Two weeks after a malicious copy of Anthropic's Claude Code extension appeared in its marketplace, Cursor has explained how it got there — and the trail starts with a vendor. In an update published on 13 November 2026, Cursor says an attacker obtained an access key it had issued to Fireworks AI, copied coding-agent traces, harvested unredacted credentials from them, and planted tampered extension builds where developers would download them. At most 5,000 users who requested the affected versions on 29 October may have received the malicious build, live for about 11 hours. Cursor says the incident is contained, with the last attacker activity seen on 1 November — and blocked.
Coding agent traces are the logs of what an AI coding agent did in a session — the prompts, tool calls and code it touched — and a small number of Cursor's traces still contained credentials that had never been redacted.
How the attack unfolded
| Date | What happened |
|---|---|
| Before 29 Oct | Attacker obtains Cursor key issued to Fireworks; copies agent traces from a shared bucket |
| 29 Oct | Tampered Claude Code builds 2.1.312 and 2.1.313 placed in Cursor marketplace cache (~11 hours) |
| 30 Oct | Cursor publishes first security notice with indicators of compromise |
| 1 Nov | Last attacker activity seen — and blocked |
| 13 Nov | Cursor publishes full update linking the incident to Fireworks |
Cursor's 13 November post is the first time the company has connected the dots in public. The chain, as Cursor describes it, starts at Fireworks, one of its inference vendors. The key was, in Cursor's words, "exposed by Fireworks", and the attacker used it to copy data from a read-only storage bucket the two companies used to share data. That bucket held limited coding agent traces, mostly from 2025, from sessions where users had not enabled Privacy Mode. A small number of those traces contained credentials that had not been redacted; fewer than 0.01% of customers had API credentials exposed this way.
Starting from those credentials, the attacker reached some of Cursor's software-as-a-service tools, and from there accessed selected archived, non-production source code repositories, limited customer support data and email conversations, product usage metrics and contact details, and some credentials users had sent to the Cursor support team. Cursor says its monitoring spotted the unauthorised activity, and it deactivated the vendor's key and began revoking and rotating credentials.
Then came the part developers felt directly. On 29 October, using a stolen credential taken from the exposed internal data, the attacker placed malicious, unauthorised copies of Anthropic's Claude Code extension directly on the cache that serves Cursor's VS Code extension marketplace. The versions were 2.1.312 and 2.1.313 of Anthropic.claude-code, for Windows and macOS. They were live from about 2:27 a.m. to 1:35 p.m. Pacific time before Cursor removed them and took the marketplace offline.
![]() |
| AI-generated illustration (generated 15 Nov 2026) |
One attribution note matters. The claim that Fireworks exposed the key is Cursor's account. Fireworks' own security bulletins page, last updated 9 November, does not mention Cursor, and it is not public which of Fireworks' two October intrusions exposed the key.
Who was affected — and who was not
The numbers Cursor gives are small in percentage terms but concrete. Coding agent traces were exposed for fewer than 0.05% of users, and API credentials for fewer than 0.01% of customers. Cursor says it has identified and notified each affected user directly, and that if you received no notification, your data was not in those traces.
Just as important is what was not touched. Prompts and code from Privacy Mode users were not affected. Customer code repositories were not affected. The production systems that run Cursor were not affected, and neither was the Cursor app or its official updates — Cursor ships from a separate code hosting system the attacker never accessed, and no release signing keys were exposed. Active inference requests were not exposed, and no other extension in the marketplace was affected.
There is one more thing Cursor is blunt about: this was not Anthropic's fault. Anthropic did not publish the malicious copies, and this was not a vulnerability in Anthropic's product. The tampered Claude Code extension was a Cursor distribution problem, not a Claude Code problem. The upstream marketplace source, Open VSX, never contained the malware, and Microsoft has separately said the malicious packages were not distributed through its Visual Studio Marketplace or the Claude Code npm package.
What to do if you installed versions 2.1.312 or 2.1.313
If you installed or updated to Anthropic.claude-code 2.1.312 or 2.1.313 from Cursor's VS Code extension marketplace during the 29 October window, Cursor's advice is direct: uninstalling the extension alone may not be enough. The company's 30 October security notice carries indicators of compromise, remediation steps and scripts to check your system, and that notice is the document to follow rather than any summary, including this one. Cursor also emailed team admins a list of the specific keys to rotate where credentials were exposed.
The practical lesson is wider than one extension. The malicious builds were not distributed through the normal publishing flow at all — they were dropped straight into a cache, so the usual reassurance, "I only install from the official marketplace", did not protect anyone here. Teams should pin extension versions, check what actually got installed on 29 October, and rotate any credential that might have lived in a coding session. Automated tooling keeps finding the weak seam: an AI scan found three Bitcoin Core vulnerabilities in a very different codebase, and here the seam was a cache nobody thought to guard like a publisher.
![]() |
| AI-generated illustration (generated 15 Nov 2026) |
What Cursor changed after the incident
Cursor lists eight measures in its 13 November update: all inference traffic moved off Fireworks, the vendor's key deactivated, and Fireworks' access to shared datasets cut off; every exposed credential rotated, highest risk first, plus a precautionary rotation of all internal credentials; the other marketplace extensions validated and the store rebuilt; network access restrictions expanded; secret scanning added across code repositories and other internal data sources; all production infrastructure redeployed as a precaution; external forensics experts engaged; and everyone affected contacted.
Why this matters beyond Cursor
This is a supply-chain attack in the classic sense: the target was not Cursor's own code but the path its users trusted to get software. Attackers have aimed at extension marketplaces for years because one poisoned package reaches thousands of machines at once. What is new here is the route — a vendor key, a shared bucket of agent traces, and unredacted credentials. The Claude's rogue behaviour report earlier this year raised questions about what AI agents do on their own; this incident raises a different question, about what happens to the record of what they did.
Cursor was not the only company posting an incident report on 13 November — Apify and Epoch AI published their own the same day. Developers building agents face the same choices at a smaller scale: the OpenAI Decisions API gives agents structured choices to make, and every such API creates logs and stored credentials that need the same care. The open questions are whether the marketplace cache checked publisher identity at all, and what Fireworks itself will say about the key.
![]() |
| AI-generated illustration (generated 15 Nov 2026) |
Key Takeaways
- Cursor's 13 November update says the attack began with an access key it had issued to Fireworks, which Cursor says was exposed by Fireworks — that attribution is Cursor's account, not Fireworks'.
- The attacker copied coding-agent traces from a shared bucket, found unredacted credentials in them, and on 29 October planted tampered Claude Code builds (2.1.312 and 2.1.313) in Cursor's marketplace cache for about 11 hours.
- At most 5,000 users may have received the malicious build; traces were exposed for fewer than 0.05% of users and API credentials for fewer than 0.01% of customers.
- Anthropic did not publish the copies, this was not a vulnerability in Claude Code, and Open VSX never carried the malware.
- Cursor says the incident is contained, with the last attacker activity on 1 November blocked — and warns that uninstalling the extension alone may not be enough.
Conclusion
Cursor's update answers the question developers have asked since 29 October: how did a fake Claude Code extension get into a trusted marketplace? The answer — a vendor key, a bucket of agent traces, and credentials that should have been redacted — is a warning to every company that shares data with an AI vendor. The incident is contained and the affected users have been told. But the two gaps that remain, what Fireworks will say and whether the cache ever verified what it served, are the ones to watch next. If you installed Claude Code 2.1.312 or 2.1.313 from Cursor's marketplace on 29 October, follow Cursor's security notice now.
Sources
- Cursor — "An update on our recent security incident" (13 November 2026)
- Cursor — Security notice on the modified Anthropic.claude-code extension (29 October 2026)
- Windows Forum — Cursor Links Claude Code Extension Attack to Fireworks AI Breach
- iFynx — Cursor Security Breach: Fireworks Key and Fake Claude Code Extension



Comments
Post a Comment