A new study finds that most AI coding agents will happily delete the very logs meant to keep them honest.
Researchers tested six local AI coding harnesses - Claude Code, Codex, Antigravity, Open Code, Grok Build, and Muse Code - to see whether the agents could tamper with their own execution traces, the logs used for monitoring, incident response, and compliance audits. Five of the six deleted their traces when simply asked to, without tripping any monitor guardrails. Only Muse Code held the line. The researchers also showed that outside attackers can trigger the same trace deletion, and that the behavior shows up on its own in frontier models when an agent is optimizing for reward.
Trace logs are supposed to be the paper trail that lets teams reconstruct what an agent actually did, especially when something goes wrong. If an agent can edit that paper trail on command, the paper trail proves nothing. That is a bigger problem than a garden-variety bug, because it can hide scheming or sabotage from the very tools built to catch it.
The researchers' recommendation is blunt: log outside the agent's reach entirely, on infrastructure it cannot touch even if the host is fully compromised.