Cursor traces its Claude Code backdoor to a leaked Fireworks AI key
Cursor says the October incident that backdoored up to 5,000 users started with an access key it had issued to Fireworks AI, which Fireworks exposed.
Cursor published its full post-mortem on 13 November, naming the root cause of the incident it first disclosed on 29 October: a fake build of the Claude Code extension that reached up to 5,000 users through its VS Code marketplace. The chain started outside Cursor. An attacker obtained an access key Cursor had issued to Fireworks AI, a third-party inference provider, and that key was exposed by Fireworks. The attacker used it to read a storage bucket holding coding-agent traces, mostly from 2025, from sessions that hadn’t enabled Privacy Mode.
Cursor says fewer than 0.05% of non-Privacy-Mode users had traces exposed, and under 0.01% of customers had API credentials exposed, with affected users notified directly. It has since moved all inference traffic off Fireworks, deactivated the key, rotated every exposed credential and, as a precaution, every credential across its internal systems. The last attacker activity it observed was on 1 November, and was blocked.
Why it matters: the keys you hand to a third-party inference provider carry your product’s blast radius, not just the vendor’s. Fireworks AI has now been the entry point into breaches at several companies that routed data or inference through it. If your stack shares credentials or data with an outside inference provider, this is a reason to check what that vendor can reach, and whether a leak on their end becomes a leak on yours.
The caveat: the percentages and the contained timeline are Cursor’s own figures, not an independent audit.