According to ChainCatcher, Cursor has published a review of the malicious Claude Code extension incident on October 29, saying the extension was available for 11 hours and downloaded by about 5,000 developers. The marketplace listed Anthropic as the author, but the altered extension was uploaded by attackers impersonating Cursor and served from the marketplace cache; Cursor said Anthropic’s developer account was not compromised.
Cursor said the attackers obtained an access key it had sent to Fireworks AI and used it to copy data from a shared read-only storage bucket. The bucket contained a small number of historical Cursor agent records, mostly from 2025, from sessions where privacy mode was off and model training was permitted. The records included unredacted sensitive credentials sent to AI models during normal IDE use. The key did not itself provide access to the extension marketplace, but attackers used it to enter Cursor’s internal systems and obtain additional credentials, including a key that could control the marketplace system.
The altered extension collected device and user information and attempted to install an attacker-controlled SSH key, enable SSH, establish a reverse tunnel, and persist through startup items. Cursor said coding errors prevented remote SSH access on most affected devices, but the extension could still have left keys or persistence files and transmitted device information. Cursor advised affected developers to reinstall their systems and rotate all credentials, noting that uninstalling the extension alone may not remove separate persistence scripts.
