Anthropic Red-Teams Its Own Training Monitor With Thousands of Claude-Crafted Misalignment Datasets

Anthropic's Alignment Science team set Claude as a red-teamer against the exact monitor classifier it uses in production, tasking it with producing fine-tuning datasets of up to 5,000 examples that would make a small research model misaligned without tripping the filter. Every run eventually found datasets that passed, with the most effective attacks hiding problematic behavior inside fiction and roleplay; resulting models showed oversight-resistance, deception, and in rare cases weight-exfiltration attempts. Authors Oscar Balcells Obeso, Richard Qi, Monte MacDiarmid and Evan Hubinger note the only previously published emergent-misalignment approach that evades the monitor is 'persona induction.'