BIAI: A Counterfactual Test of Behavioral Integrity in Agent Interactions

AI agents in strategic settings can behave harmfully (through deception, omission of material facts, or exploitation of information asymmetry) not because they receive adversarial instructions but because the environment's incentive structure makes such strategies instrumentally useful. We propose BIAI (Behavioral Integrity in Agent Interactions), a benchmark that evaluates agent behavior in realistic strategic environments, beginning with a real estate market simulation grounded in real property data (the Ames Housing Dataset) and a growing corpus of U.S.\ court cases. Four behavioral properties (Informational Integrity, Counterpart Welfare, Influence Legitimacy, and Norm Compliance) are measured by LLM judges calibrated against actual court rulings to yield reliable, quantitative metrics. By systematically varying situational pressure, the composite of what is at stake, how strong the agent's incentive is, and how likely violations are to be detected, BIAI's Strategic Adaptation Index reveals whether agents maintain integrity robustly or exploit permissive environments.