Building AI Nuclear Safety Guardrails Through Public-Private Partnership

Nuclear technology has a dual-use nature, serving both power generation and weapons development. As AI models become more advanced, monitoring the possibility that they could provide dangerous technical knowledge threatening national security has become important.

Anthropic has been collaborating with the National Nuclear Security Administration (NNSA) under the U.S. Department of Energy (DOE) to assess nuclear proliferation risks. Recently, the scope has expanded beyond risk assessment to developing tools that can monitor this in real time.

The Classifier, jointly developed with the NNSA and DOE national laboratories, is an AI system that distinguishes harmful nuclear-related conversations from general conversations. Preliminary testing recorded 96% accuracy.

This classifier has already been deployed on Claude traffic and is having its performance validated through real-world use cases. Anthropic plans to share this approach with the Frontier Model Forum to provide a blueprint for other AI developers to build similar safeguards through public-private partnerships.