Back to News
News · AI Agents & Security

Cursor Confirms Breach Led to a Malicious Claude Code Extension

A stolen credential let attackers publish a tampered copy of Anthropic's extension on Cursor's marketplace.

Source: BleepingComputer · November 14, 2026

BleepingComputer reports that Cursor has confirmed a breach at Fireworks AI led to a malicious copy of Anthropic's Claude Code extension being distributed through Cursor's own VS Code marketplace. Cursor's account is that attackers obtained an access key held by Fireworks AI and used it to reach a storage bucket shared between the two companies, which held historical AI coding-agent traces, some of which contained unredacted API credentials that let the attackers reach additional Cursor systems. On October 29, the attackers used a stolen credential to publish the tampered extension, inserting a malicious JavaScript loader at the start of the extension's code while leaving the rest of it unchanged. Anthropic's own extension and the Open VSX registry were not affected.

Cursor estimates that up to 5,000 users who requested the affected versions during the exposure window may have received the malicious build, while it says fewer than 0.05% of users had coding-agent traces exposed and fewer than 0.01% had API credentials compromised. Coding errors in the payload likely stopped the attackers from gaining remote access to many of the affected Windows and macOS machines. Fireworks AI had already disclosed multiple security breaches in October, including one involving exposed API keys and tokens.

Why it matters: AI coding agents are now a supply-chain attack surface in their own right, a single compromised partner credential let attackers push a trojanized extension under a trusted developer's name. Enterprises should pin and verify the source of every AI coding tool and extension, rotate credentials after any partner breach, and treat AI-agent trace data as sensitive, since it can carry embedded secrets.