VEPA Uses Verified Policy Evolution to Guard LLM Agents

Summary

LLM agents increasingly act on private documents and mutable system state, creating safety and security risks when actions are irreversible. VEPA is a framework for evolving formal runtime guardrails as new risks appear during deployment. It represents guardrails as executable symbolic rules over typed agent events, rather than relying on an LLM to interpret a policy at every decision point. An LLM uses deployment feedback to propose patches for newly observed failures, but a deterministic verifier accepts only structurally valid, conflict-free policies that preserve previously correct behavior. An evidence-aware authority layer then activates a verified rule only after enough supporting evidence has accumulated, preventing uncertain updates from directly changing decisions. The paper reports consistent improvements over baselines on ASSEBench-Safety, ASSEBench-Security, and ATBench, while broadening deterministic formal-policy enforcement. The reported gains also persist under mixed event streams, corrupted feedback, and longer deployment horizons. The abstract does not provide the numerical margins or further details of the experimental setup.