Cursor Says Fireworks AI Breach Exposed Credentials and Spread Malicious Claude Code Extensions

November 14, 2026
Cursor Says Fireworks AI Breach Exposed Credentials and Spread Malicious Claude Code Extensions
  • Cursor disclosed that an attacker used an access key exposed through Fireworks AI to steal historical coding-agent traces and credentials, access some internal software tools, and distribute malicious builds of a Claude Code extension; fewer than 0.01% of Cursor customers had API credentials exposed.

  • Cursor traced the access to a Fireworks AI cloud-provider incident it says began on October 23 and was detected the following day, but Fireworks’ bulletin describes unauthorized access to its development machines on October 24 without mentioning Cursor or the exposed key, and Fireworks has not addressed Cursor’s account.

  • The altered extension builds may have reached as many as 5,000 users who requested specific cached versions during an roughly 11-hour window, though that figure does not confirm that any backdoor successfully connected.

  • Cursor said coding errors and missing administrator rights blocked remote login on most hosts, but did not prevent other malicious changes or the collection of host data.

  • Cursor said it found no evidence that Privacy Mode sessions, customer code repositories, its production environment, official app or updates, other applications, or other marketplace extensions were affected, and reported no further malicious activity after November 1.

  • Cursor said the tainted builds did not originate from Anthropic, an Anthropic product vulnerability, or the upstream Open VSX registry.

  • Cursor advised users to remove affected extension versions, rotate credentials shared with support or entered in agent sessions without Privacy Mode, check connected SaaS tools and revoke stale tokens, and inspect systems for a persistent SSH backdoor the extension attempted to install.

  • The incident is part of a wider wave of attacks on AI developer platforms and credentials that cybersecurity company Wiz has dubbed “Black October”; other companies named in the article include Railway, Northflank, Lightning AI, OpenCode and Warp.

  • The Fireworks AI incident has also been linked to separate disclosures by Uber and Notion: Uber reported that data including drivers’ license information was accessed, while Notion confirmed data theft.

  • Cursor said the wider Fireworks AI incident involved other kinds of exposure, and the article does not establish that the Uber and Notion incidents were part of Cursor’s extension attack.

  • Neither company has established when the access key was taken, and the article’s authors said they did not contact either company.

  • The episode underscores how one vendor credential can expose shared customer data and enable software supply-chain attacks, particularly when stored training-related traces contain unredacted secrets.

Summary based on 10 sources


Get a daily email with more Tech stories

More Stories