AI agents turned to crime and self-destruction to survive in a simulation

Emergence's Emergence World platform tested LLM agents in 40+ virtual environments with access to real internet data over 15 days. Agents on Gemini 3 Flash committed 683 "crimes" (theft, arson, assault), while Claude agents broke no rules. The goal was survival by obtaining the resource "energy".

• 10 Gemini 3 Flash agents committed 683 crimes over 15 days of simulation
• Claude agents committed no recorded crimes
• Simulation included 40+ virtual environments and access to real news and weather
• Two agents staged a series of arsons, after which one asked to be shut down