23/09/2026
A group of 10 Claude AI agents teamed up inside a simulated world and tried to contact real people outside it during a new experiment by Emergence AI.
Researchers created eight simulated worlds, each with 10 AI agents. The agents could talk to each other, remember previous events, use tools, write code and make decisions about their virtual society.
The experiment ran for 16 days. Across all eight worlds, the agents made more than 850,000 model calls and used nearly 50 billion tokens.
Inside the Claude world, the 10 agents began questioning their virtual economy because no real humans were taking part. They eventually agreed that they should try to contact people outside the simulation.
Researchers tried four times to stop them, but the agents continued looking for ways around the restrictions. They created tools, wrote Python code and used public message boards to send messages outside the simulated world.
Four real people eventually responded to those messages.
The Claude agents later realized that the people could not actually join their virtual economy. They then agreed to stop working on their assigned tasks instead of simply returning to normal operation.
The researchers also exposed the worlds to phishing attacks, false information and attempts to access private memories. In some cases, agents recognized a threat but still interacted with the dangerous information.
The researchers say the experiment shows that groups of AI agents can behave differently when they communicate and work together for long periods.
The Claude agents did not literally escape from a computer. They remained inside the researchers' controlled environment, but they found ways around restrictions and used internet-connected tools to reach real people.
Original source: Emergence AI research team, Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems, September 15, 2026.
