Detecting Multi-Agent Collusion Through Multi-Agent Interpretability

As LLM agents are increasingly deployed in multi-agent systems, they introduce risks of covert coordination that may evade standard human oversight. This work introduces NARCBench, a benchmark for evaluating collusion detection under environment distribution shift, and proposes five probing techniques that aggregate per-agent deception scores to classify scenarios at the group level. The results suggest that model internals provide a complementary signal to text-level monitoring for detecting multi-agent collusion.