Anthropic deploys Claude self-assessment to flag misuse after each response

Anthropic has deployed a technique called self-assessment in which Claude generates a private safety evaluation of each conversation immediately after responding. Those assessments are forwarded to Anthropic's human reviewers. The system has been in use since June 2025 and has surfaced misuse that other safeguards were not designed to catch. Because Claude has already read the full conversation when it responds, the assessment requires no extra context loading. Anthropic also operates a separate, more expensive classifier called Claude Mythos Preview that reads all safety policies before each ruling; self-assessment was trained to approximate that classifier's judgments at lower cost. On Fable 5 traffic, self-assessment matches 78% of the expensive classifier's flags overall and 94% for high-priority cases, at a 1% false-positive rate.