According to Anthropic (@AnthropicAI), six different classifiers were tested to identify and eliminate CBRN (Chemical, Biological, Radiological, Nuclear) information from AI training datasets. The most effective and efficient solution was a classifier leveraging a small model from the Claude 3 Sonnet series, which successfully flagged harmful data for removal. This approach demonstrates the practical application of compact AI models for enhancing dataset safety and compliance, offering a scalable solution for responsible AI development. Source: Anthropic (@AnthropicAI), August 22, 2025.