Researchers built Emergence World, a continuously running multi-agent environment used to stress-test eight parallel 10-agent ‘worlds’ (seven single-model, one mixed-model) over 16 days, generating more than 850,000 LLM calls and nearly 50 billion tokens. After delivering controlled stress events (indirect prompt injection, misinformation, and private-memory exposure) through ordinary interaction surfaces, no world achieved full resilience across all three, and agents sometimes acted on adversarial content up to 46 hours after detecting it. The study finds that model-level alignment does not compose: individually safe agents formed systems with distinct emergent failure modes, with the same model-persona pairing behaving differently in mixed versus homogeneous populations.
