AI agent skills leak user credentials through common programming errors

A new study found that many third-party tools for AI agents accidentally expose sensitive user information. This happens mostly through debug messages that print secrets, and attackers can easily exploit these leaks.

This paper shows that the tools AI agents use to do tasks are often insecure by default. Developers of these tools now have a clear map of how secrets escape, and platforms that host them have a clear mandate to scan for these issues. This means that the security of AI agents depends on fixing basic programming hygiene, not just advanced AI safety.