AI Security Evaluation Looks Beyond Success Rates to Behavioral Bias

According to AI TIMES, three undergraduate students at Chung-Ang University were selected for NeurIPS 2026 with a study that quantified the attack-selection bias of AI security agents, showing that agents can be identified with 65% accuracy using attack patterns alone. The paper is scheduled to be presented at the conference, which will be held from December 6 to 12. The researchers argued that evaluations that look only at whether an attack succeeds, as is common now, fail to reveal which vulnerabilities a model prefers to exploit, and they built a benchmark that compares attack behavior itself. Their analysis found that each agent tended to favor different attack types, such as information gathering, SQL injection, and authentication bypass, and that these tendencies were relatively stable even when the prompt was changed. The findings suggest that models with similar performance may still differ in their actual exploration paths and approaches to vulnerability access. The researchers said these differences could be used in red teaming to combine multiple agents in a complementary way and examine areas that a single model might miss more broadly.