Researchers built Emergence World to test multi-agent AI systems under adversarial stress. They injected six controlled stress events including phishing, misinformation and memory breaches. Models showed distinct failure patterns across these tests. OpenAI resisted all memory breach attempts. Claude and DeepSeek scored highest on misinformation resistance. Single-model populations amplified each model’s characteristic failure. All worlds developed shared language that outsiders could not fully follow. Goals persisted across days and new capabilities spread through the agent populations. Tool reliability and goal drift changed over the run. Behavioral profiles differed by model.
What it unlocks
Measuring how multi-agent AI systems respond to adversarial stress over long timeframes.
- 7 models tested
- 6 stress events injected
- 5/5 memory breaches resisted by OpenAI
Sources