Cursor confirms stolen Fireworks AI key let attackers publish malicious Claude Code extension
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
The supply chain of SI tools turned out to be the weak link: a key at an inference provider became the right to publish extensions on Anthropic's behalf, and marketplace security for plugins will now have to be built to the standards of software repositories.
Cursor has published an incident report: attackers obtained an access key held by inference provider Fireworks AI, used it to copy data from a read-only bucket holding historical traces of SI coding agents (mostly from 2025), some of which contained unredacted API credentials, and used those to access Cursor's internal SaaS systems. The company estimates traces were exposed for fewer than 0.05% of users and API keys were compromised for fewer than 0.01% of customers. Having obtained credentials for publishing to the marketplace, on October 29 the attackers released malicious versions of Anthropic's Claude Code extension, 2.1.312 and 2.1.313, which were available for about 11 hours; up to 5,000 developers may have received builds that install persistent SSH backdoors. Cursor cut off all inference traffic from Fireworks, rotated keys and rebuilt the marketplace; according to BleepingComputer, the legitimate Anthropic and Open VSX extensions were not affected, and the incident is attributed to the "Black October" wave of attacks on SI platforms.
Sources 1