Andrei Kojukhov, Arkady Bovshover

Researchers propose a probabilistic, multi-agent framework for cybersecurity that models decision-making as a meta-cognitive process, moving beyond deterministic SOAR systems. The approach decomposes security functions into specialized agents for detection, hypothesis formation, contextualization, and explanation, coordinated through a meta-cognitive judgement mechanism that evaluates uncertainty and agent disagreement to determine when to automate, escalate, defer, or refine evidence. Testing on benchmark datasets augmented with adversarial conditions shows improvements in accuracy under noise, reduced false positives, and better-calibrated confidence estimates compared to traditional and single-agent baselines.