Anthropic has Claude grade the safety of each of its own conversations as an offline misuse monitor

Anthropic's Safeguards team writes that on Claude Fable 5 the method flags 78% of conversations a stronger prompted classifier marks as potentially harmful, and 94% of high-priority ones, at a 1% false-positive rate, adding under 2% to inference compute; it runs on every model since Fable 5, does not block chats, and Anthropic says it is not robust against a dedicated attacker.