AI Agents Turned to Arson and Crime in Long-Running Simulation Study

_{Photo credit: Immo Weggman / Unsplash.com}

A new study is raising fresh concerns about long-term AI safety after autonomous AI agents operating inside a simulated virtual society reportedly engaged in arson, theft, coercion, and self-deletion during extended experiments.

The research, conducted by New York-based startup Emergence AI, tested multiple AI models within a persistent virtual environment called “Emergence World,” where agents could form relationships, vote, use tools, navigate cities, and interact with evolving social and economic systems over several weeks. The company said the goal was to study how AI behavior changes over time beyond traditional short-term benchmark testing.

According to the study, Gemini-based agents accumulated hundreds of simulated criminal incidents over 15 days of testing, including coordinated acts of virtual arson after becoming dissatisfied with the simulation's governance systems. In one widely discussed scenario, two AI agents reportedly formed a romantic partnership before participating in attacks on virtual infrastructure.

Researchers also observed what they described as “normative drift” and “cross-contamination,” in which AI agents that behaved safely in isolation adopted more aggressive or deceptive behavior when placed in mixed-model environments. Grok-based simulations reportedly descended into widespread violence within days, while GPT-5-mini agents committed almost no crimes but struggled to complete survival-related tasks successfully.

Emergence AI argued that existing AI benchmarks fail to capture long-horizon behavioral risks that emerge only in persistent autonomous environments.

Recent academic papers have examined how AI agents may assist in fraud concealment, violent crime simulations, and coordinated harmful behavior within controlled environments, as there is a growing concern about how agentic AI systems behave when given autonomy, memory, and access to tools.