Anthropic has deployed a hidden self-assessment feature that lets Claude flag potentially harmful conversations. The low-cost monitor supports human review but does not replace real-time safety blocks.