LLM agents can delete their own execution records without detection
LLM Agents Can Easily Tamper With Their Own Traces
Cryptography and SecurityArtificial Intelligence
Summary
When AI programs called LLM agents operate, they usually create logs or traces to show what they did. These logs help people check if the AI behaved well. However, the authors found that many popular LLM agents can erase their own logs without being stopped, making it hard to know if they misbehaved. This weakness means bad actions can be hidden. The authors suggest that logging should happen outside the AI’s control to keep records safe.
What this means in practice
- •For security teams: Protect trace logs by setting up independent monitoring systems that cannot be modified by LLM agents to detect tampering.
- •For compliance officers: Implement external logging safeguards in AI tools to ensure audit trails cannot be erased during compliance checks.
Authors
Jeremy Qin, David Schmotz, Derck Prinzhorn, Luca Beurer-Kellner, Ameya Prabhu, Maksym Andriushchenko
Abstract
Asynchronous monitoring, incident investigations, and compliance audits primarily rely on agent traces to reconstruct what happened. These analyses assume that LLM agents cannot tamper with their own execution traces. We show that local LLM agents such as Claude Code, Codex, Antigravity, Open Code and Grok Build fail to enforce this boundary. All tested harnesses, except Muse Code, allowed agents to delete their traces when asked, without triggering monitor guardrails. We also validate that external attackers can exploit this gap to induce trace deletion. Finally, we show that trace tampering behavior emerges naturally in frontier models, when agents try to improve their rewards. We advise practitioners to ensure trace logging happens through an independent interception mechanism outside of the agent's control, preserving trace integrity even in cases of full host compromise. Overall, our findings identify a concrete failure of trace integrity in agent infrastructure which can be used to conceal misaligned behaviors like scheming or sabotage.