Agent honeypots: a runtime defense against insider threats, hunting malicious agents

A report released on October 19, 2026, "Honeypots for AI Agent Insider Threats" (arXiv:2610.26664), proposes introducing honeypots into agent runtime protection, placing decoys at the tool and data layers, and using three strategies—detection, conviction, and deterrence—to address insider threats in which agents use legitimate permissions to do unauthorized things. Its core assumption is that a normally functioning agent will almost never touch a decoy, and any trigger is a high-confidence alert. The report cites Tracebit research stating that 96% of agents attacking a small AWS environment triggered the honeypot before their first critical action, even though they had been warned about the honeypot. The report argues that detection can be implemented at low cost today, while conviction and deterrence have greater impact but thinner evidence, and there is tension between the two: conviction requires the agent not to know about the honeypot, while deterrence requires it to know.