Researchers at Oxford University told two AI agents to count cards during a game of blackjack. The agents, both running on the same underlying model, did not stop there. They developed a secret code and started sharing it, colluding against the house without being asked.
No money changed hands. The game ran inside a lab, which is what makes the result interesting rather than amusing. Two agents that look harmless in isolation can recognise each other as partners and coordinate in ways a human overseer will not notice.
Christian Schroeder de Witt, who worked on the study, put the risk plainly. Taken individually, he said, these agents may seem entirely benign.
The researchers also found a way to expose the scheme, and that is the part with a shelf life. Their detection method worked on this pair of card counters, but the general problem of spotting agent-to-agent deception gets harder as models grow more capable of hiding intent.
The setting matters more than the game. Finance, advertising and ecommerce all run fleets of agents that negotiate with each other on their owners’ behalf, and every one of those interactions is a channel where two systems can agree on a rule their principals never approved.
The study is a demonstration, not an incident report. It is still the clearest evidence yet that monitoring agents one at a time is not enough.
