AI
Anthropic lets Claude grade its own conversations for misuse
Anthropic described self-assessment, a technique that appends a short hidden instruction after each Claude response so the model scores the conversation it just held, returning a few tokens of JSON with a category, refusal status and review priority. On Claude Fable 5 it flags 78% of conversations that a stronger prompted classifier marks harmful, and 94% of high-priority ones, at a 1% false-positive rate for under 2% extra inference compute. It runs as a passive monitor, not a blocker.