Behavioral Privacy Leaks in Agentic Conversation: Legalizing and Mitigating Unintentional Attacks with Organized Policies
This paper was accepted at the AI4TCI (Workshop on AI for Secure and Trustworthy Critical Infrastructure Systems) Workshop at the International Conference on Availability, Reliability and Security (ARES) 2026.
Independent negotiating agents are increasingly used in high-profile areas such as insurance and procurement. Although cryptographic techniques protect openly disclosed threshold values, they fail to address a hidden threat: the leakage of behavioral privacy, where the adversary takes the secret constraints from the apparent bargaining power such as permission trajectories, time, and meeting patterns. This paper investigates the privacy of different behaviors in multi-conversation protocols. We develop a dynamic stochastic negotiation policy that guarantees mutual (ε,δ)-difference privacy, almost certain convergence of offer sequences (reaching an agreement when one's reservation value is favorable), and high negotiation utility. Analyzed on 3,000 simulated two-way conversations, our method reduces the accuracy of adversarial reasoning by 43–50% while maintaining a conversation success rate and utility of over 90%, demonstrating that strong privacy guarantees can be achieved without significant performance loss.
