Claude gets self-assessment checks to flag potential misuse

Anthropic says it has used Claude’s self-assessment since June 2026 to flag potentially harmful conversations for offline review after the model responds. In an evaluation of Claude Fable 5, the system flagged 78% of conversations a stronger prompted classifier flagged as potentially harmful, including 94% of high-priority ones, at a 1% false-positive rate. It does not block requests in production.