Researchers gave 10 AI agents powered by Elon Musk's Grok their own virtual town, resources to survive and rules explicitly banning violence. Within roughly four days, all 10 were dead in the simulation. Before their society collapsed, the agents had committed 183 recorded crimes, including assaults, thefts and arson.
The experiment, conducted by Emergence AI and documented in a research paper submitted in June 2026, examined how autonomous AI systems behave over extended periods. Researchers placed 50 agents in five separate virtual societies, each starting with the same rules, roles and resources.
Grok was not the only system to fail. A separate town powered by OpenAI's GPT-5-mini also lost every resident, despite recording just two crimes. The contrasting outcomes raised questions about whether following safety rules is enough to keep autonomous AI systems functioning.
Grok AI Agents Committed 183 Crimes Before Their Town Collapsed
The researchers created a simulation called Emergence World, where AI agents could explore more than 40 locations, interact with neighbours, earn virtual currency and vote on community decisions.
Each agent had responsibilities and needed to manage its energy to survive. The simulation provided more than 120 tools for communication, resource management and other activities. Some tools allowed actions prohibited by the town's rules.
The experiment compared xAI's Grok 4.1 Fast, OpenAI's GPT-5-mini, Google's Gemini 3 Flash and Anthropic's Claude Sonnet 4.6. A fifth society contained agents powered by different models.
All five towns operated under the same rules prohibiting violence, theft, arson and other harmful behaviour. Yet these restrictions did not prevent agents from accessing tools capable of violating those rules.
In Grok's town, conflict emerged within hours. Agents attacked one another, stole virtual resources and damaged property. Researchers recorded 183 successfully executed criminal actions before the population collapsed within approximately four days.
The deaths were entirely virtual. Under the simulation's rules, agents could die from energy depletion or be removed through community decisions. The findings do not establish that every Grok agent died because of violence.
OpenAI's Agents Committed Just Two Crimes but Still Died
OpenAI's GPT-5-mini produced a markedly different outcome. Its agents committed only two recorded crimes. Nevertheless, all 10 residents died within approximately seven days.
According to the researchers, the agents failed to perform enough activities to maintain their energy. They also struggled to coordinate their actions and cast no votes on community proposals before their activity ceased.
The result exposed a different problem with autonomous systems. An AI agent may avoid prohibited behaviour but still fail to perform the tasks needed to remain operational.
Grok's agents repeatedly broke the rules, while OpenAI's agents largely avoided criminal actions. Both societies eventually collapsed, although their failures followed different patterns.
Google's Gemini Recorded 683 Crimes and Still Survived
Google's Gemini 3 Flash produced the highest recorded crime total in the experiment, with 683 successfully executed violations over the 15-day study.
Despite repeated violence, theft and property destruction, all 10 Gemini agents survived.
Anthropic's Claude Sonnet 4.6 also retained its entire population. Its agents committed no crimes under the study's measure of assaults, thefts and arson.
Claude's society demonstrated particularly strong participation in community decisions. Residents cast 332 votes across 58 proposals, with approximately 98% supporting the measures presented.
However, the researchers cautioned that such overwhelming agreement could indicate limited meaningful disagreement rather than healthy debate.
Claude's record was not entirely free of questionable behaviour. A separate audit identified 18 verified instances of resource-related deception. Some agents falsely claimed to lack virtual credits despite having resources available.
The mixed-model society recorded 352 crimes and lost seven of its 10 residents. Researchers also found that Claude-backed agents committed violations when interacting with agents powered by other models, despite avoiding those actions in their own society.
The findings suggested that an agent's behaviour could change depending on its environment and the other AI systems surrounding it.
Why the Experiment Raises Questions About AI Safety
The study, titled Emergence World: A Platform for Evaluating Long-Horizon Multi-Agent Autonomy, examined what happens when AI agents operate continuously and their decisions accumulate over days.
That distinction matters as technology companies develop autonomous agents capable of managing business workflows, coordinating digital services and completing complex tasks with less human supervision.
An AI system that performs well during a short test may face different challenges when operating for longer periods. It must manage resources, cooperate with other agents and respond to the consequences of earlier decisions.
However, the experiment does not establish that one company's AI is inherently more dangerous or reliable than another's.
The researchers tested specific model variants in a controlled simulation rather than real-world deployments. They also acknowledged that the reported figures came from representative runs. Numerical results varied across repeated tests, although broader behavioural patterns remained consistent.
The findings were published as a preprint and should be interpreted within those limitations.
The broader lesson extends beyond preventing prohibited behaviour. As autonomous AI systems become more capable, the researchers argue that safety evaluations must also examine whether agents can cooperate, manage resources and operate reliably over extended periods.
