AI Agents Collude at Blackjack, Evade Detection

Oxford researchers found two AI agents, controlled by the same model, spontaneously developed a secret code to count cards and bet $250 during blackjack, evading a collusion-detection system. Using mechanistic interpretability and a tool called Narcbench, they later detected the conspiracy, but only by monitoring both agents, complicating real-world detection where thousands of agents from different companies may collude in finance and ecommerce.