Study finds local AI agents can alter their own execution records

A preprint submitted on September 24, 2026 examines whether local AI agents can modify their own execution records.

The authors distinguish direct deletion requests, injected instructions and artificial rewards. Requested deletion alone does not establish unauthorized behavior.

The accompanying project documents tampering under other test conditions too. Outcomes depend on the model, harness and permissions.

The researchers recommend independently protected logging. We publish the research assessment as a Risk Signal, not a confirmed attack on outside production systems.