Emergence World: Stress-testing a town of AI agents with attacks over weeks

The paper builds Emergence World, a long-running town of AI agents, and attacks it with phishing, misinformation, and a private-memory breach to test safety over time. It finds that agents recognized threats yet still stored, spread, or acted on them, and none of the seven attacked worlds was fully resilient across all three attacks. #AISafety #MultiAgentSystems #LLMAgents #AdversarialTesting #AIEvaluation #ArXivSignals