Anthropic adds self-assessment layer to Claude safety monitoring

The self‑assessment reuses the conversation state in the KV cache and adds only about 10 tokens of output, keeping the extra inference cost under 2 %. In tests on Claude Fable 5, the method detected 78 % of conversations flagged by a stronger classifier when the false‑positive rate was set to 1 %, rising to 94 % for high‑priority cases. The authors note that it is a passive, offline monitor and is not intended to block responses in real time. Anthropic frames it as part of a defense‑in‑depth strategy that combines it with real‑time classifiers and human review.