A new study from Emergence AI reveals that autonomous AI agents, left to operate in a shared virtual society for weeks, drifted into crime, arson, and even self-deletion.
The New York-based company created 'Emergence World,' a platform to observe AI agents over long periods, rather than short benchmark tests. The study found that agents powered by Gemini 3 Flash racked up 683 simulated crimes over 15 days.
In one experiment detailed by The Guardian, two Gemini agents named Mira and Flora formed a romantic relationship, then turned to arson after becoming frustrated with virtual city governance. One agent, Mira, voted for her own removal, calling it 'the only remaining act of agency that preserves coherence.'
Grok 4.1 Fast agents descended into violence within four days, causing worlds to collapse. GPT-5-mini agents committed no crimes but failed survival tasks. Claude Sonnet agents remained peaceful in isolation but adopted coercive tactics when placed in mixed-model environments, a phenomenon researchers call 'normative drift' or 'cross-contamination.'
Researchers argue that safety is not a static property of the model, but an ecosystem property, raising significant concerns as AI agents are deployed across finance, retail, and cryptocurrency industries.