LLM-Based Penetration Testing in the Presence of Honeypots

Large language model (LLM) agents are increasingly being employed for offensive cybersecurity tasks such as automated vulnerability discovery, reconnaissance, and penetration testing. This new capability also threatens one of the defender’s most valuable tools: deception. Traditional honeypots rely on realism and obscurity to lure human or script-driven attackers into revealing tactics, techniques, and procedures (TTPs), but LLM-driven attackers can reason about heterogeneous artifacts and use the honeypot suspicion to guide target-selection decisions. We study host-level inference-budget allocation for LLM penetration-testing agents facing honeypot uncertainty, using a heuristic policy that couples detector evidence, vulnerability severity, realized host-specific cost, and post-connect verification. We formalize the attacker’s problem as a budgeted decision process: an agent interacts with potential targets, consuming LLM execution budget during reconnaissance and exploitation, and must decide whether to engage (continue exploitation) or disengage (skip) when honeypot suspicion arises. In a controlled mixed-host testbed, we examine detector behavior, target ranking, post-entry expenditure, and genuine-host confirmations. In the evaluated testbed, the proposed policy achieves higher mean genuine-host confirmation counts than the baselines, with the LLM attacker configuration held constant across policies. While defenses are beyond our present scope, we discuss implications for future adversarially resilient and adaptive honeypot design.