Reproduce “Efficient Auditing of Adversarial AI Agent Behavior from Agent Traces” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Eugene Zhang, Cheng-Yun King Yang, Dongyan Xu

The authors present selective auditing: deterministic rules choose pending actions for an LLM auditor to examine with preceding trace context. Development failures jointly refine gate rules and audit instructions. On OpenAgentSafety, tuned gating achieves 72.8% TPR and 5.9% FPR with 2.33 audits and 14.6k tokens per run, but detects fewer unsafe traces than auditing every action.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.