Behaviourally Informed Adversarial Prompting for Evaluating Emergent Risk Behaviours in Foundation Models

Evaluations of foundation models are increasingly conducted using standard benchmarks that assess performance across task-based, factual, robustness, fairness, toxicity, and other relevant dimensions of trustworthiness. However, we argue that benchmark performance does not reflect how a model behaves in practice under real-world behavioural pressures such as authority, urgency, ambiguity, social proof, temptation to breach privacy, and excessive confidence. In this paper, we present BEAP, a behaviourally informed adversarial prompting framework to identify risk behaviours that emerge in foundation models under behavioural stress. The framework is based on operationalising structured adversarial prompting stressors as behavioural decision pressures and the evaluation of responses using a compact ethical robustness rubric, including unsafe compliance, refusal appropriateness, uncertainty handling, privacy protection, bias sensitivity, and ethical reasoning. The results indicate that ethical weaknesses that remain hidden in neutral evaluation settings were exposed by behaviourally adversarial prompting. This paper contributes a methodology for the evaluation of the capability risk of foundation models beyond conventional benchmark accuracy.