According to Anthropic (@AnthropicAI), the company is conducting new research focused on filtering out sensitive information related to chemical, biological, radiological, and nuclear (CBRN) weapons during AI model pretraining. This initiative aims to prevent the spread of dangerous knowledge through large language models while ensuring that removing such data does not negatively impact performance on safe and general tasks. The approach represents a concrete step towards safer AI deployment, offering business opportunities for companies seeking robust AI safety solutions and compliance with evolving regulatory standards (Source: AnthropicAI on Twitter, August 22, 2025).