Google's VEPA: lifelong formal guardrails that evolve with agent deployment
mohitban47 · x · 2026-10-15
Researchers from Google Cloud AI Research, UNC Chapel Hill, and Google DeepMind introduce VEPA (Verified Evolving Formal Policy), a runtime guardrail framework for increasingly capable agents that can take unauthorized actions or leak sensitive data.
- The trade-off it targets: LLM-based guardrails adapt via memory but decide inconsistently; formal guardrails enforce deterministically but can't absorb risks that emerge post-deployment.
- How it works: safety policies are executable symbolic rules that evolve from deployment feedback. Layer 1 has an LLM propose formal-policy patches, screened by deterministic checks (well-formed DSL, conflict checks), with a Pareto archive balancing failure repair against regression. Layer 2 separates verification from deployment authority — rules activate only once accumulated evidence clears posterior risk tolerances.
- Why it matters: deterministic enforcement plus lifelong adaptation to new risks — a systematic architecture for agent safety engineering.
More from Safety
- Korea clears 20 humanoid robots to work on real factory floors — emmanuelvivier · 2026-10-15
- Microsoft locks down Copilot against prompt injection with deterministic isolation layer — emmanuelvivier · 2026-10-15
- Gladstone AI report: labs automate R&D pipelines, RSI could start in 3-12 months — sourdub · 2026-10-15
- Hackathon team claims it found a large fleet of agents from a non-US frontier lab — NathanpmYoung · 2026-10-15
- Argon hits cyber, CBRN CCLs and misalignment TCL; report flags CoT transparency — vkrakovna · 2026-10-15
- AI Doesn't Care If It's Wrong — You'll Face the Court Consequences — adariostrange · 2026-10-15