Published
- 4 min read
By Allan D - Editor, AI Security Wire
Cursor Says a Key Exposed at Fireworks Led to the Fake Claude Code Extension
Update to our earlier coverage of the trojanized Claude Code extension and the Fireworks AI breach. In a post dated November 13, 2026, Cursor gives its account of how the fake extension got onto its marketplace, and it points to Fireworks AI, a vendor it shared data with. Cursor describes it as the picture its investigation “pieced together”.
This is Cursor’s account. As of 15 Nov 2026 we found no confirmation from Fireworks, Anthropic or an independent party in the sources we checked. We checked the Cursor posts and two Fireworks bulletins (the 9 October original and the November update). Neither Fireworks bulletin mentions Cursor.
The chain, as Cursor tells it
Cursor says an attacker obtained an access key it had issued to Fireworks, and that Fireworks exposed the key. The attacker used it to copy data from a read-only storage bucket Cursor used to share data with the vendor.
The bucket held limited coding agent traces, mostly from 2025, from sessions not in Privacy Mode, which Cursor says it was permitted to use for model training. A small number contained credentials that had not been redacted. Cursor says less than 0.01% of customers had API credentials exposed.
Cursor says that starting with those credentials, the attacker reached select SaaS tools it uses, and through them archived non-production source code repositories, limited support data and emails, product usage metrics, summaries of some coding agent conversations and some credentials users had sent to support.
Separately, Cursor says that on October 29 the attacker, “using a stolen credential from the exposed data from our internal system”, placed malicious copies of the Claude Code extension directly on the cache that serves its marketplace. Cursor does not say which route that credential took.
What Cursor says about the extension
Cursor says Anthropic “did not publish this copy, and this was not a vulnerability in Anthropic’s product”, and that Open VSX, the marketplace’s source, “never had the malware”. It says the build was available for about 11 hours and that at most 5,000 users who requested specific cached versions on October 29 potentially received it.
Our October 31 article, citing BleepingComputer, gave the same 5,000 figure; we have not independently checked that outlet. The affected builds are Anthropic.claude-code 2.1.312 and 2.1.313 for Windows and macOS, per Cursor’s October notice, whose remediation steps still apply. Cursor warns that uninstalling the extension alone may not be enough.
Scope and status, per Cursor
- Privacy Mode prompts and code: not affected.
- Live traffic and production systems: not affected. Cursor says the exposed trace data was historical storage used for offline analysis.
- The Cursor app and official updates: Cursor says it ships from a separate code hosting system the attacker never accessed, and no release signing keys were exposed.
- Other marketplace extensions: Cursor says it validated them all.
- Traces for users not in Privacy Mode: exposed for fewer than 0.05% of users, who Cursor says it notified. It says that if you got no notification, your data was not in these traces. We cannot check that.
Cursor says it believes the incident is contained, with the last attacker activity on November 1, blocked. It says it moved all inference off Fireworks, deactivated the key, rotated every exposed credential and then all internal credentials as a precaution, rebuilt the marketplace, restricted network access so stolen credentials cannot be used outside their intended location, and added secret scanning. It also says it temporarily disabled internal and product services well beyond what the evidence required.
Fireworks says its evidence indicates links to a financially motivated group (it calls the practice LLMjacking) and describes a third-party cloud provider breach. Cursor attributes nothing, and neither source establishes a link between the two.
Why it matters
Our assessment: on Cursor’s account, this began with a key and a data-sharing arrangement, not a flaw in a model or an extension. Coding agent sessions hold pasted tokens, and a copy of those sessions in a training or analysis bucket carries the secrets with it. If you share agent logs or traces with a vendor, check what credentials they could contain, scope the vendor’s keys to the bucket and network location they need, and redact before sharing.
References
Frequently Asked Questions
- How does Cursor say the fake Claude Code extension reached its marketplace?
- Cursor says an attacker obtained an access key it had issued to Fireworks, which Fireworks exposed, and used it to copy data from a read-only bucket. On October 29, it says, the attacker used a stolen credential from the exposed data to place malicious copies of the extension on the cache that serves its marketplace. This is Cursor's account only.
- How many people were affected?
- Cursor says at most 5,000 users who requested specific cached versions of the extension on October 29 potentially received the malicious build. It says coding agent traces were exposed for fewer than 0.05% of users not in Privacy Mode, and that less than 0.01% of customers had API credentials exposed. Cursor says it contacted each affected customer directly.
- What does Cursor say about Anthropic and Open VSX?
- Cursor says Anthropic "did not publish this copy, and this was not a vulnerability in Anthropic's product", and that the marketplace source, Open VSX, "never had the malware". Those are Cursor's statements. As of Nov. 15, 2026 we found no Anthropic or Open VSX statement in the sources we checked.