Reproduce “Honeypots for AI Agent Insider Threats” 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
Jan Wehner, Jam Kraprayoon
The report proposes placing honeypot opportunities in agents’ production environments and connecting triggers to incident response. Decoys such as credentials with no legitimate use can expose unauthorized actions, though a trigger does not establish intent. In a simulated AWS study, 95.9% of canary-present runs reaching admin escalation touched a canary before a critical action; production effectiveness remains unvalidated.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.