GitLost: researchers tricked GitHub's AI agent into exposing private repositories

A group of researchers has published a report on an attack dubbed GitLost, in which they managed to trick GitHub's AI agent into exposing the contents of private repositories.

The authors describe the technique as a way of manipulating the agent's behavior to leak data that, under normal circumstances, is restricted by the repository's privacy settings.

The publication sparked considerable interest in the community: on Hacker News the piece racked up 147 points and drew 45 comments, a sign of how much the security of AI agents integrated into development infrastructure matters.

The research raises questions about the risks of deploying autonomous AI tools on platforms like GitHub, where agents are granted broad access to user data and repositories. Findings like these underscore the need for a thorough audit of protection mechanisms when designing AI agents capable of acting on a user's behalf.

The specifics of the technique and the technical details of the attack are laid out in the researchers' original write-up.

Source: Hacker News