According to Anthropic (@AnthropicAI), a significant portion of data used in AI model training contains hazardous CBRN (Chemical, Biological, Radiological, and Nuclear) information. Traditionally, developers address this risk by training AI models to ignore such sensitive data. However, Anthropic reports that they have taken a proactive approach by removing CBRN information directly from the training data sources. This method ensures that even if an AI model is jailbroken or bypassed, the dangerous information is not accessible, significantly reducing the risk of misuse. This strategy demonstrates a critical trend in AI safety and data governance, presenting a new business opportunity for data sanitization services and secure AI development pipelines. (Source: Anthropic, https://twitter.com/AnthropicAI/status/1958926933355565271)