Out-of-Band Policy Enforcement Secures AI Agent Tool Access

Summary

The paper presents Out-of-Band Policy Enforcement (OBPE), a trusted boundary outside an AI agent’s reasoning. It limits typed operations and resources, narrows queries, filters or masks responses, and can block requests using semantic conditions. In 3,621 trials involving four models and adaptive red-team tasks, failures involving protected data or forbidden effects fell from 57.6% to 0.2%. Safe-and-useful completion improved, although fulfillment declined and the authors identify reconstruction and side-channel leakage risks that remain unresolved.