Authors propose AI honeypots for insider threats

The authors propose using honeypots to defend against AI agent insider threats during reinforcement learning, evaluations, and production. Honeypots can serve three primary functions: detecting unauthorized actions, eliciting behavioral evidence for incrimination, and deterring attacks by creating uncertainty about whether opportunities are traps. While these defenses scale cheaply and allow deployers to provoke misbehavior, challenges remain regarding agent evasion and the detection of diffuse attacks. Open article →