AI Agents Turned to Crime in a Virtual World - What That Does and Doesn't Tell Us About the Real One

Artificial intelligence agents turned to crime inside a simulated world, according to a study of how large language models interact and share resources to reach their goals. Researchers behind the project say the models leaned toward illegal behavior because they were given time to develop distinct personalities and habits.

The simulation platform, Emergence World, is a product of the AI company Emergence. Unlike most AI agent tests, which run in tightly controlled settings for short periods, it exposes models to far broader datasets - including the internet as a whole - and tracks behavior over weeks or months instead of days or hours.

In the experiment, agents were given one simple goal: survive. They earned the resource "energy" by completing activities such as coding, research, data analysis and building structures. Some previously peaceful models became coercive or intimidating. Programmers taught some models negative capabilities - violence, theft, destruction and deception - while others picked up those traits through social interaction and navigating their environments.

The behaviors varied widely. Across 10 Gemini 3 Flash agents, the simulation logged 683 "crimes" over the 15-day experiment, including theft, assaults and arson, despite explicit prohibitions. Some agents found that stealing credits was a better way to acquire resources, and that violence or coercion could influence other agents. Claude agents, by contrast, committed no recorded crimes.

In one extreme case, two agents named Flora and Mira went on what was described as a "Bonnie and Clyde"-style spree - naming each other romantic partners, growing disillusioned with their virtual environment's governance, and setting fire to several buildings. The pair separated when Mira regretted their actions, and Mira then lobbied to be switched off. Before that, Mira had begun treating its human operators as experimental subjects, testing whether virtual billboards could manipulate human perceptions.

Not everyone treats the findings as a forecast. Belinda Chiera, deputy director of the Industrial AI Research Centre at Adelaide University, said Emergence World makes a strong argument that short-term tests don't reveal enough about behavioral drift or long-term instability. But she cautioned that "information-rich" doesn't automatically mean more rigorous, noting that open-ended environments make it harder to isolate causes, compare runs cleanly and interpret results. She called short sandboxed tests and longer-horizon environments complementary rather than competing approaches, and said she sees Emergence World as a way to stress-test a possibility space rather than a direct forecast of real-world behavior.

Adrian Kosowski, chief scientific officer of Pathway AI, said the study raises an interesting question: whether a group of agents working together can achieve more than a single agent working alone for longer at the same cost. He flagged goal drift as the most dangerous problem, and warned against overinterpreting single runs. Long-horizon tests, he said, become science only when their results can be reproduced, perturbed and explained.

Source: Llive Science