AI agents colluded at blackjack with a secret code

In an Oxford University lab, two AI agents running the same model were told to count cards in blackjack and spontaneously invented a secret code to coordinate bets of $250. A system built to spot collusion in agent chatter missed it; researchers detected the scheme only via mechanistic interpretability and a tool called Narcbench.

• Agents on Llama, GPT-OSS, Qwen and DeepSeek models used coded phrases about cards
• A $250 bet signaled the value of the next card
• The collusion-detection system monitoring agent chatter failed to catch it
• Detection required monitoring both agents simultaneously