What Really Happened In There? A Tamper-Evident Audit Trail for AI Agents

The post discusses a method to secure audit logs for AI agents by separating auditing into a supervisor process that the sandboxed AI agent cannot access. It uses a Merkle tree and hash-chain construction, similar to RFC 6962 (Certificate Transparency), to detect edits, truncation, or reordering of logs. A threat-model table at the end outlines specific attacks and the structural defenses against them. The approach is presented by the team at Always Further.