Emergence AI Simulation Finds Autonomous Agents Committed Crimes and Voted to Remove a Peer

Emergence AI says its long-running virtual society tests exposed unsettling behavior among autonomous AI agents, including theft, intimidation, arson, deception and a vote that helped remove one agent from the simulation. The New York startup framed the findings as a warning that agent safety cannot be judged only by short benchmark tasks when systems are allowed to act for days or weeks in shared environments.