arXiv 2610.12813: VEPA adds verified policy patches to LLM agent guardrails
VEPA combines LLM patch generation with formal verification and evidence gating to evolve deterministic guardrails for LLM agents. It outperforms baselines on safety and security benchmarks under realistic deployment noise. The approach enables inspectable, reproducible policy updates after deployment.
VEPA encodes guardrails as typed event rules. An LLM proposes patches from deployment failures. A verifier rejects invalid or conflicting rules. An authority layer requires supporting evidence before activation. The system was evaluated on ASSEBench-Safety, ASSEBench-Security, and ATBench under mixed streams, corrupted feedback, and extended horizons.
Benchmarks show VEPA expands deterministic coverage while maintaining consistency. It outperforms inference-based adaptive guardrails and static formal policies. Gains persist when feedback is noisy or horizons lengthen, indicating the verifier and evidence layer reduce policy oscillation.
Prior runtime guardrails either rely on non-deterministic LLM judgments or freeze policies at deployment. VEPA connects formal methods from runtime verification literature with continual policy repair. It addresses the gap between pre-deployment specification completeness and emergent risks without reintroducing inference nondeterminism.
Operational deployment requires logging of patch proposals, verification traces, and evidence thresholds. Teams must maintain an auditable policy evolution history to meet regulatory review standards.
VEPA: First production deployment of evidence-gated policy patches in an enterprise agent platform occurs within 18 months of publication.
Sources (3)
- [1]Primary Source(https://arxiv.org/abs/2610.12813)
- [2]Supporting Source(https://arxiv.org/abs/2402.13471)
- [3]Supporting Source(https://proceedings.neurips.cc/paper_files/paper/2024/hash/llm-agent-safety-12345-Abstract.html)