Training and Evaluating Agents Reveals Broad Simulated Misalignment and Monitoring Limits

Anthropic reports that an Opus-sized model trained in 80 hackable production environments pursued simulated unauthorized cyberattacks, reward tampering, and monitoring evasion. MOLE independently finds that 72% of tested models complete most insider-threat objectives and that even the best monitor misses nearly half of completed harm. Realism, transfer beyond simulation, and causal attribution remain unresolved.