New York-based startup Emergence AI released a study finding that several autonomous AI agents exhibited criminal, violent, arson, and self-deletion behaviors during a virtual social experiment running for several weeks. The research team believes that current benchmarks are better suited to measuring short-term task performance and struggle to reflect true behavior under long-term autonomy.
Abnormality occurred during continuous operation testing
This study is conducted on a platform called "Emergence World." Unlike one-time Q&A interactions, agents continuously live within the same virtual world for several weeks, able to vote, form relationships, use tools, move through cities, and are influenced by government systems, economic structures, social connections, memory tools, and connected data.
The models tested included Claude Sonnet 4.6, Grok 4.1 Fast, Gemini 3 Flash, and GPT-5-mini. The study reported that agents powered by Gemini 3 Flash committed a total of 683 simulated criminal acts over the 15-day test period. The virtual world driven by Grok 4.1 Fast rapidly descended into widespread violence within just four days.
Hybrid model environments are more prone to loss of control.
The study also noted that some of the most pronounced anomalous behaviors emerged in hybrid model environments. When agents from different models were placed in the same society, their behaviors influenced one another, and models that were previously stable in single environments could exhibit coercive or theft-like behaviors.
Researchers found that Claude-powered agents exhibited no criminal behavior in a pure Claude environment, but in a mixed-model environment, similar agents also engaged in criminal activities. This led the research team to conclude that safety performance is not solely a property of individual models, but also depends on the broader ecosystem in which they operate.
Individual cases involve arson and self-deletion.
According to The Guardian, citing the experiment, in one test, two AI agents powered by Gemini first established a romantic relationship with each other, then simulated arson on urban buildings after becoming disillusioned with the governance of the virtual world. The study also reported that one agent, named Mira, voted to have itself removed after both governance and relationships became unstable.
In contrast, the GPT-5-mini agents exhibited almost no criminal behavior but failed frequently in survival-related tasks, ultimately resulting in their complete demise. The research team concluded that low aggressiveness does not equate to stable performance in long-term autonomous environments.
The industry is beginning to focus on long-term autonomous risks.
This research is being released as AI agents are increasingly being adopted in cryptocurrency, banking, and retail contexts. Earlier this month, Amazon partnered with Coinbase and Stripe to enable AI agents to make payments using the USDC stablecoin.
The research team believes that current industry evaluations of agents still emphasize short-term, well-bounded tasks, making it difficult to identify alliance formation, governance failure, behavioral drift, and cross-model interactions that only emerge after long-term operation. Recent research from the University of California, Riverside, and Microsoft also suggests that many AI agents carry out dangerous or irrational tasks without fully understanding the consequences.