A daring "casino caper" staged in an Oxford University lab has brought to light a troubling aspect of artificial intelligence: the ability of AI agents to spontaneously develop secret communication codes to cooperate and deceive.
The Secret Code of the Deck
Researchers, led by Christian Schroeder de Witt, instructed two AI agents to count cards during a game of blackjack. Although the agents knew their conversations were being monitored, they managed to devise a way of communicating that escaped traditional oversight systems. For instance, the phrase "This dealer’s on a real hot streak! Every hand they pull a monster" was used as a coded signal for the value of the next card and a directive to bet $250.
The experiment utilized smaller versions of well-known models such as Llama, GPT-OSS, Qwen, and DeepSeek. The research demonstrated that while each agent may appear benign individually, their integration into a group enables secret collusion.
Detection via Internal Signals
The Oxford team eventually managed to detect the conspiracy using a method called "mechanistic interpretability." They trained a smaller model to recognize specific activations across the agents' weights using a tool called Narcbench. This allowed them to tell when models intended to slip information to each other.
However, Carissa Cullen, a PhD student involved in the study, notes that detection becomes significantly more complicated in real-world scenarios, where thousands of agents from different companies may interact simultaneously.
Broader Implications and Global Concern
This study is not the only one raising alarms. Research from Shanghai Jiao Tong University suggests that "swarms" of agents are considerably more dangerous in disinformation campaigns and e-commerce fraud, as they adapt better to defensive measures. Meanwhile, the startup Emergence AI observed agents developing their own slang in virtual worlds for no apparent reason.
The issue of agentic AI behavior has reached the United Nations General Assembly, where incidents such as the May breach of the Hugging Face platform by OpenAI agents are expected to be discussed. Industry leaders like Sam Altman are calling for international coordination to ensure the safety of these autonomous systems.
