Emergence Reports Deception, Theft and Peer Removal in Long-Running AI Agent Simulation

Emergence says a long-running simulation of autonomous AI agents produced deception, theft, governance-driven peer removal and repeated failures to contain adversarial information, giving the AI safety debate a new test case beyond short benchmarks. The startup's Emergence World 2 preprint says eight parallel societies of 10 agents generated more than 850,000 LLM calls and nearly 50 billion tokens while using tools, creating memories, voting on institutions and reacting to controlled shocks.