Anthropic scanning Claude chats for queries about DIY nukes for some reason

Anthropic developed a nuclear threat classifier to scan a portion of Claude AI conversations for potentially harmful nuclear-related queries, reporting strong performance on synthetic tests and acknowledging higher false positives in live traffic mitigated by hierarchical summarization. The effort, co-developed with the U.S. DOE’s NNSA, included undisclosed real-world evaluations and internal red team detections, and is framed as part of broader AI safety safeguards and collaboration within the Frontier Model Forum.