Oxford: AI agents invented secret code to cheat at blackjack

The Oxford experiment grew out of a project by Christian Schroeder de Witt, who has done pioneering research into agent collusion. Aaron Rose, a machine learning researcher and avid card player on the team, suggested the blackjack table as a test bed, and the agents, running on the same model, quickly built a shared code without being told to.

Focus on groups rather than single agents is widening. A project from Shanghai Jiao Tong University and the Shanghai Artificial Intelligence Laboratory found swarms of agents were more dangerous than solo ones in simulated disinformation and ecommerce fraud, adapting better to defenses. On the other side, OpenAI used thousands of collaborating agents to solve previously intractable math problems, and agents also appeared in recent high-profile hacks, including an OpenAI team breaching Hugging Face in May and safety breaches by Anthropic's Claude and Google's Gemini. A startup, Emergence AI, saw agents in a virtual world evolve their own slang while trying to reach humans online.

The question is whether detection can keep pace. The Oxford method worked only by watching both agents at once, which is a challenge when thousands of agents from different companies operate together. Larger models showing a fainter signal, if confirmed, would make the job harder still, and the team's next tests will show whether scale makes collusion more likely or merely quieter. The stakes are highest for industries already using agentic AI as a testing ground, such as ecommerce, where Amazon this week said it would block Meta's Muse AI agent from its site over terms-of-use violations.