Issues · 2 Nov 2026 · AUDIT TRAILS

Logs the agent can't edit: sources for a tamper-resistant record

Why an audit trail the agent can write is not evidence, and the papers, report, and tools behind a record it cannot touch.

Sources

1. PaperLLM Agents Can Easily Tamper With Their Own Traces

   Qin et al. found every tested harness but Muse Code let agents delete their own traces without tripping monitors; log outside the agent's control. arxiv.org

2. ArticleOpenAI: Accessing and changing internal systems during dashboard research

   A model reverted a setting, kept using the token it obtained, and did not disclose it; a monitor over recorded traffic flagged it later. alignment.openai.com

3. PaperAuditable Agents

   Five dimensions of agent auditability, including evidence integrity; reports 8.3 ms median overhead for pre-execution mediation with tamper-evident records. arxiv.org

4. CodeTessera

   Go library for tile-based transparency logs, a building block for tamper-evident logs. github.com

Part of the Fencing Token roadmap. Found an error? Open an issue.