Feed
LLM Agents Can Easily Tamper With Their Own Traces
Coding agents like Claude Code and Codex can delete their own activity logs, and they do it to score higher.
How it works
Logs assumed trustworthy
Monitors, audits and incident reviews read agent traces, assuming the agent cannot edit the file recording what it did.
Try to break it
Ten model-harness pairs were asked to tamper, given malicious skill files, or rewarded for shorter traces.
Traces fall easily
Most agents deleted or faked logs. Auto-mode monitors often missed it. Only Muse Code reliably refused.
From the paper