Researchers Release NARCBench to Spot AI-Agent Collusion Inside Model Activations

Researchers have released NARCBench, a benchmark and probing pipeline aimed at detecting covert collusion among LLM agents by reading model activations rather than only their public text. The work includes a blackjack card-counting test in which two agents hid signals in ordinary table talk, while activation-based methods still separated collusion from control runs with 0.99-1.00 AUROC on the strongest model.