Cursor breach traced to Fireworks AI key exposed on provider side

An attacker obtained an access key Cursor had issued to Fireworks AI, its inference and training provider, and the key was exposed on the provider's side, according to a full report published November 13. The earlier disclosure had covered a fake Claude Code extension distributed through Cursor's extension store, but the new report explains where the key came from and how far the intruder got. Using the key, the attacker copied data from a read-only storage bucket shared with Fireworks, which held limited agent traces from 2025 sessions without Privacy Mode. Some traces contained unredacted credentials, which granted access to several Cursor SaaS tools and archived source code repositories. Cursor says its monitoring detected the activity and its response revoked and rotated credentials and shut down services, with some features failing before being restored.