Anthropic Has a Plan to Keep Its AI From Building a Nuclear Weapon. Will It Work? (www.wired.com)

Anthropic says it worked with the Department of Energy and the NNSA to ensure its chatbot Claude won’t help build a nuclear weapon by deploying frontier versions of Claude inside AWS “Top Secret” cloud enclaves where NNSA teams red‑teamed the models. Over months of iterative testing they co‑developed a “nuclear classifier” — essentially a filter built from an NNSA-curated list of nuclear risk indicators and sensitive technical topics (the list is controlled but not classified) — tuned to catch concerning conversations while avoiding false positives for benign nuclear energy or medical isotope discussions. Anthropic is offering the classifier to other AI firms and frames it as a proactive, shared safety measure. The effort is significant because it shows a practical, hands‑on model for government—industry collaboration on AI risks, and it moves beyond vague promises to deploy concrete tooling and secure testing environments. But experts warn the approach has limits: if Claude never contained classified nuclear material to begin with the tests are less revealing; classification secrecy prevents independent evaluation; and future model capabilities or subtle LLM failure modes (e.g., bad math or synthesis of disparate open research) could still pose risks. The classifier is a useful guardrail, but not a definitive solution—broader transparency, continual red‑teaming, and industry adoption will determine how effective it is at preventing genuine misuse.