Anthropic Uses Claude Self-Assessments to Flag Potential Misuse

Anthropic says Claude now assesses conversations for potential misuse immediately after responding, adding a monitoring signal for further review. The company reports that the system has surfaced activity missed by other detection methods since its production rollout in June. The technique, which Anthropic calls self-assessment, asks Claude to produce a compact safety assessment of the full conversation, including requests it refused. The assessment can identify a potential harm category, assign a priority for human review, and indicate whether Claude refused. Anthropic says the assessments feed offline review and help identify possible misuse that existing classifiers may miss. (Anthropic’s technical overview)