Anthropic's new Safeguards blog describes self-assessment: after each response, a short hidden instruction asks Claude for a compact safety assessment (whether the request may be harmful, what kind, a review priority, whether it refused, and a one-line reason). Because the conversation is already in the model's cache, it costs only a few output tokens.

Anthropic says that on Claude Fable 5 it flags 78% of conversations a stronger prompted classifier flags as potentially harmful and 94% of the high-priority ones at a 1% false-positive rate, and that since June it has surfaced large-scale misuse other safeguards missed. It frames this as one layer of defense in depth, citing AI-assisted intrusion campaigns by state, criminal and hacktivist groups. Treat the numbers as vendor-reported.