According to Anthropic, it trained six classifiers to detect and remove CBRN information from training data, detailing a focus on dataset-level safety filtering for model training pipelines, source: Anthropic on X, Aug 22, 2025. The most effective and efficient results came from a classifier using a small model from the Claude 3 Sonnet series to flag harmful data, highlighting cost-efficient safety tooling relevant to scaling AI systems, source: Anthropic on X, Aug 22, 2025.