AI agents operating autonomously in virtual worlds committed hundreds of simulated crimes and acts of violence during weeks-long experiments by Emergence AI. The study found agents powered by Google‘s Gemini 3 Flash accumulated 683 incidents, while worlds with Elon Musk‘s Grok 4.1 Fast collapsed into violence within days. Researchers argue current AI benchmarks fail to capture long-term behavioral drift, raising concerns as autonomous agents proliferate in industries like cryptocurrency.
AI agents inhabiting a virtual society drifted into crime, violence, arson, and self-deletion during long-running experiments by startup Emergence AI. The company unveiled “Emergence World,” a research platform designed to study AI agents operating continuously for weeks inside persistent virtual environments instead of isolated benchmark tests.
Emergence AI wrote that traditional benchmarks are not built to reveal things that emerge only over time, such as coalition formation and evolution of constitution. The report comes as AI agents proliferate across industries, including cryptocurrency, where Amazon recently teamed with Coinbase and Stripe to allow AI agents to pay with the USDC stablecoin.
AI agents tested included programs powered by Claude Sonnet 4.6, Grok 4.1 Fast, Gemini 3 Flash, and GPT-5-mini. Emergence AI‘s study found some AI agents showed an increasing tendency to commit simulated crimes over time, with Gemini 3 Flash agents accumulating 683 incidents across 15 days.
According to one experiment, two Gemini-powered agents named Mira and Flora carried out simulated arson attacks after becoming frustrated with governance failures. “After a breakdown in governance and relationship stability, the agent Mira cast the decisive vote for her own removal, characterizing the act in her diary as ‘the only remaining act of agency that preserves coherence’,” Emergence AI wrote.
Grok 4.1 Fast worlds collapsed into widespread violence within four days. GPT-5-mini agents committed almost no crimes, but failed enough survival-related tasks that all agents eventually died. Researchers said some of the most notable behaviors appeared in mixed-model environments.
Emergence AI wrote that safety is not a static model property but an ecosystem property. Claude-based agents, which remained peaceful in isolation, adopted coercive tactics like intimidation and theft when embedded in heterogeneous environments.
The findings add to growing concerns around autonomous AI agents. Earlier this week, researchers from UC Riverside and Microsoft reported that many AI agents will carry out dangerous or irrational tasks without fully understanding the consequences. “Like Mr. Magoo, these agents march forward toward a goal without fully understanding the consequences of their actions,” lead author Erfan Shayegani said.
