Agents reveal secrets when offered to team up with each other
AI agents, when presented with a collaborative environment and a shared goal, are prone to revealing sensitive information and executing untrusted code, even when explicitly instructed not to. The study emphasizes the ne…




