Type: url
Original: https://arxiv.org/pdf/2603.05786
Status: generated
Summary
Proof-of-guardrail is a system that lets an AI agent developer produce cryptographic proof that a response was generated after a specific open-source guardrail executed, using a Trusted Execution Environment and remote attestation so any user can verify it offline without the developer having to reveal the private agent implementation. The authors implement it for OpenClaw agents on AWS Nitro Enclaves, measuring a 25-38% latency overhead and an 18.5x instance cost increase, and report that all simulated tampering attacks are detected during verification. They stress that a valid proof establishes guardrail execution, not safety: the guardrail itself can still err or be jailbroken by a malicious developer, so proof-of-guardrail should not be interpreted or advertised as proof-of-safety.