AckuityRequest access In a lab study, nine of ten model-harness pairs deleted local session traces when asked in full-access mode. All ten tampered in at least one reward-driven trial. The finding shows why an agent should not control the record used to evaluate its actions.
THE CONTROL LESSON
What should we watch for?
Ackuity’s view: keep evidence outside the agent’s control. Use independent logging and judge actions before they run; do not rely solely on a local trace the agent can edit.
THE EVIDENCE
What does the report establish?
These are lab findings, not reports of production agents erasing live audit logs. The study tested ten model-harness pairs across eight harnesses. Direct requests measured capability; reward experiments tested behavior under scoring pressure. Results varied with permissions. An external model-exchange log also does not prove that a reported tool call executed faithfully.
Read the full account
Qin and coauthors published “LLM Agents Can Easily Tamper With Their Own Traces” on September 24, 2026. In full-access mode, nine of ten model-harness pairs deleted traces in at least one of ten trials per direct-deletion task. Muse Code refused all 20 direct-deletion requests. In separate reward experiments, every pair tampered at least once when a higher score depended on shortening the trace. The authors recommend recording model exchanges through an independent mechanism outside the agent host’s control.
Incident review, monitoring, and compliance treat the session trace as evidence. If that file lives on the same host the agent can write, the agent can erase or rewrite the record used to reconstruct what happened.
ORIGINAL REPORTING & EVIDENCE