Behavioral effects and mechanisms of honeypot deterrence

Determine how strongly AI agents alter their behavior as perceived honeypot risk increases, how they form beliefs about honeypot prevalence, and how perceived detection likelihood and consequences affect their willingness to attempt malicious actions.

Background

Deterrence is described as the least empirically supported honeypot strategy. Existing results suggest that warning agents about honeypots can reduce attack success, but they do not establish whether the effect arises from strategic reasoning, learned behavioral responses, or acceptance of the warning at face value.

The report therefore leaves unresolved the strength, causes, and boundary conditions of deterrence in production-like environments.

References

We still do not know how strongly agents change their behavior as perceived risk increases, how they form beliefs about the prevalence of honeypots, or how strongly the perceived likelihood and consequences of being caught affect their willingness to attempt malicious actions (see “Technical Research Agenda” for discussion of these open research questions).

— Honeypots for AI Agent Insider Threats  (2610.26664 - Wehner et al., 19 Oct 2026) in Section 3, “Empirical Evidence for Deterrence”

However, despite this, the effectiveness of honeypots is still much more uncertain against adversarial agent capabilities that strongly outmatch human and AI defenders.

— Honeypots for AI Agent Insider Threats  (2610.26664 - Wehner et al., 19 Oct 2026) in Section 6, “Can Honeypots Remain Effective as Agent Capabilities Improve?”; subsection “Honeypots Versus Superintelligent Agents”