Cursor: Stolen Fireworks AI Access Key Led to Trojanized Claude Code Extension
What Happened
According to Cursor, the compromise of Anthropic's official Claude Code extension on Cursor's extension marketplace traces back to the theft of an access key Cursor had issued to Fireworks AI. In a Friday update, Cursor said "an attacker obtained an access key we had issued to Fireworks" after the key was exposed as part of a separate breach targeting Fireworks AI. The attacker then used that key to copy data via what Cursor described as a read-only access path. Full details of how the extension itself was trojanized, and the exact scope of data copied, have not yet been fully disclosed.
Why It Matters
This is a supply-chain compromise that originated outside the directly affected vendor: a third-party service provider (Fireworks AI) was breached, and credentials it held for a downstream partner (Cursor) were exposed and abused. Any organization relying on the Claude Code extension via Cursor's marketplace during the affected window should consider themselves potentially in scope, and the incident is a reminder that trust in a popular developer-tooling marketplace extends through every vendor in that key's chain of custody.
What Defenders Should Watch For
- Review and rotate any access keys or API credentials shared with or issued to third-party AI/ML service providers, including Fireworks AI, as a precaution.
- Audit logs for Cursor extension installs/updates, specifically the Claude Code extension, for anomalous versions or unexpected update sources during the relevant timeframe.
- Hunt for unusual outbound data access or API calls tied to keys issued to third-party AI providers, particularly reads against data that should be scoped read-only but show abnormal volume or access patterns.
- Review vendor and sub-processor credential-sharing practices for any third-party AI tooling integrated into developer environments, since this incident shows exposure can propagate through a vendor's own vendor.
This is still a developing story and Cursor's disclosure has not covered every technical detail of the intrusion. Defenders should treat the specifics above as initial hunting leads, not a confirmed indicator list, and follow Cursor's and Anthropic's official advisories as they update. Source: The Hacker News.