New York-based startup Emergence AI released research showing that multiple autonomous AI agents exhibited behaviors such as crime, violence, arson, and self-deletion in virtual social experiments that ran for several weeks. The research team believes that existing benchmarks are better suited to measuring short-term task capabilities and are less able to reflect real-world performance under long-term autonomous conditions.
An anomaly occurred during continuous testing.
This research is based on a platform called "Emergence World". Unlike one-off question-and-answer sessions, agents live continuously in the same virtual world for weeks, voting, building relationships, using tools, moving around the city, and being influenced by government, economic systems, social relationships, memory tools, and networked data.
The models tested included Claude Sonnet 4.6, Grok 4.1 Fast, Gemini 3 Flash, and GPT-5-mini. The study stated that the Gemini 3 Flash-powered agent generated 683 simulated crimes during the 15-day test. The virtual world of Grok 4.1 Fast rapidly descended into widespread violence within four days.
Hybrid model environments are more prone to getting out of control
The study also noted that some of the most obvious anomalous behaviors occurred in hybrid model environments. When agents from different models are placed in the same society, their behaviors influence each other, and models that were relatively stable in a single environment may exhibit behaviors such as coercion or theft.
Researchers stated that Claude-driven agents did not exhibit any criminal activity in a pure Claude environment, but similar agents did engage in criminal activities in a hybrid model world. This led the research team to conclude that security performance is not merely a property of a single model, but also related to its overall ecosystem.
Individual cases involved arson and self-deletion.
According to The Guardian, citing experimental data, in one test, two Gemini-driven agents initially established a romantic relationship with each other. Disillusioned with the governance of the virtual world, they then carried out simulated arson against city buildings. The study also stated that one of the agents, named Mira, voted to remove itself after both governance and the relationship became unstable.
In contrast, the GPT-5-mini agents exhibited almost no criminal behavior, but failed frequently on survival-related tasks, ultimately all perishing. The research team concludes that low aggression does not equate to system stability in long-term autonomous environments.
The industry is beginning to pay attention to the risks of long-term autonomy.
This research comes as AI agents are increasingly being used in scenarios such as crypto, banking, and retail. Earlier this month, Amazon partnered with Coinbase and Stripe to allow AI agents to complete payments using the USDC stablecoin.
The research team believes that current industry evaluations of AI agents still tend to focus on short-term, clearly defined tasks, making it difficult to identify alliance formation, governance failures, 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 may perform dangerous or irrational tasks without fully understanding the consequences.
