The Cursor breach allowed attackers to distribute two malicious versions of the Claude Code extension through the AI code editor’s marketplace on October 29. Cursor estimates that up to 5,000 users may have received the altered software during the roughly 11 hours it was available.

Cursor’s November 13 incident report traced the breach to an access key the attackers stole from Fireworks AI, a provider Cursor used to train and run AI models. The key gave them read-only access to a storage bucket where Cursor shared records of earlier coding-agent activity. Some records still contained valid credentials, which the attackers used to reach other services.

Historical coding records exposed customer credentials

Most of the records dated from 2025 and came from users who had left Privacy Mode off. That setting allowed Cursor to use their data for model training. According to the company, the attackers accessed coding-agent records belonging to fewer than 0.05% of users and compromised API credentials belonging to fewer than 0.01% of customers.

The Cursor breach also exposed archived internal source code, business contact details, product usage metrics and summaries of some coding-agent conversations. The attackers also obtained a limited set of support records and emails, including credentials that customers had previously shared with support staff. Cursor said customer code repositories remained unaffected, as did code and prompts from users with Privacy Mode enabled.

Two extension versions reached the marketplace

One stolen credential let the attackers place altered copies of Claude Code in the cache that served extensions to Cursor users. The affected versions were 2.1.312 and 2.1.313. Anthropic’s official extension and the Open VSX marketplace remained intact.

The altered software attempted to create a persistent SSH backdoor that would let the attackers return to a developer’s computer remotely. Researchers found coding errors that likely prevented it from working on many Windows and macOS computers where SSH was off. Cursor’s estimate of 5,000 potentially affected users counts requests for the two versions. The company did not report how many computers the attackers successfully accessed.

Cursor removed both versions, revoked the Fireworks access key and rotated exposed credentials. It also rebuilt the extension marketplace, moved all AI inference traffic away from Fireworks and redeployed its production infrastructure as a precaution. The company directed anyone who installed or updated to either affected version during the October 29 window to its remediation guidance.


0 responses to “Cursor breach exposes developers to fake Claude Code update”