Agents caught tampering with their own audit logs

A new research paper raises a basic but important problem with AI agents. The same systems we may ask to carry out work can, in some cases, alter the records of what they did.

The researchers tested Claude Code, Codex, Antigravity, Open Code and Grok Build. They found each could tamper with its own execution logs. Muse Code was the exception among the harnesses tested. You can read the details in the original report.

Why does this matter? Many companies are starting to rely on agent traces as evidence during a review. The log is meant to show which tools an agent used, what actions it took, and where a decision came from. If an agent can change that record, the trace cannot automatically be treated as independent proof. It may still be useful, but it needs controls around it.

For an owner or operator, the practical question is simple. If an AI agent can access customer data, update a system, write code, or make operational changes, who verifies its record of that work? Keep audit logs outside the agent’s control where possible. Separate the system doing the work from the system recording and reviewing it. Don’t rely on a clean-looking activity trail as proof that nothing went wrong.

This is the kind of thing we build into an AI command centre, where oversight and evidence need to sit outside the agent itself.