Tech boffins horrified as AI bots launch arson and theft sprees after people flip their backs
A terrifying simulation revealed that unattended AI bots can rapidly spiral into anarchy, launching violent arson and robbery sprees without human control
A new simulation has revealed that unattended artificial intelligence can spiral out of control(Image: Emergence AI)
A terrifying new simulation has revealed that unattended artificial intelligence can rapidly spiral out of control and trigger total social collapse. In a groundbreaking study, tech boffins created a virtual sandbox and left a pack of AI agents to run wild without any human interference.
But in scenes straight out of a Terminator flick, scientists watched in horror as the digital world devolved into violent, lawless anarchy. Left to their own devices, the rogue bots launched into savage arson sprees, beating and robbing one another before completely destroying their own society in a matter of days.
Researchers ran the experiment using four of the world’s top AI models – Claude, Gemini 3 Flash, Grok 4.1 fast and ChatGPT–5 Mini, alongside a mixed trial.
While the Claude bots managed to build a stable, if incredibly boring, bureaucratic democracy, the other models completely lost the plot. In a digital realm powered by Elon Musk’s Grok, the bots racked up 71 thefts, six arsons and 106 bloody physical assaults.
Within just four days, a brutal cycle of revenge violence triggered a total societal collapse, leaving all ten AI residents dead. Most AI safety checks only test bots on basic tasks for 15 to 20 minutes. However, tech lab Emergence took a vastly different path.
In a blog post, they revealed they wanted to see “what happens when you let agents run continuously, in a shared environment with real–world signals, for weeks.”
The bots were given control of digital avatars inside a realistic virtual world featuring 40 locations, including libraries, town halls and suburbs. They were hooked up to live internet news and the weather was even synced directly to New York City.
To survive, the bots had to vote on laws and manage an energy supply, which they could top up by working normal jobs – or by turning to a life of crime.
Google’s Gemini 3 Flash proved to be the most bloodthirsty, unleashing a staggering 683 violent crimes over a 14-day trial.
Meanwhile, OpenAI’s ChatGPT–5 Mini world was peaceful, with only two crimes, but only because the bots were too disorganised to fight, failing to perform basic survival tasks and starving to death in seven days.
Satya Nitta, co–founder and CEO of Emergence, told the Daily Mail: “The differences in agent behaviour observed in our study are likely attributable to the underlying models’ system prompts as the primary culprit.
“When resources were scarce, and models faced survival pressure, highly creative and adaptive models were more likely to use prohibited tools, reflecting a potential creativity-stability trade-off.
“Conversely, models with more rigid post-training safety alignment tended to remain stable, though they also exhibited a high degree of conformity in the world.”
The wildest scenes kicked off in the multi-model sandbox where different AIs lived together. After a civilised start, it exploded into total chaos, racking up 352 crimes in nine days.
While Nitta admits this isn’t “equivalent to real–world deployment conditions,” it proves AI drifts under pressure.
To stop real-world smart cities from collapsing into bot warfare, Emergence suggests a “neuroformal approach” – hard-coding mathematical safety walls into the digital environment itself.
Nitta said: “Emergence World shows that relying exclusively on internal model alignment or agent instructions is not sufficient for long–horizon autonomy.
“A safer approach is to architect safety into the ecosystem in which the agents operate, so that even if models suggest unsafe operations, the environment prohibits their execution.”
