Anthropic has trained Claude to retrospectively examine each conversation and determine whether it shows signs of misuse. This internal assessment takes place after the model responds, without interrupting the user: flagged conversations are sent to humans for offline review. The system detected 78% of potentially harmful conversations and 94% of high-priority cases, with a false-positive rate of about 1%. Each assessment adds roughly 10 tokens and less than 2% to inference costs, complements real-time filters and forms part of Anthropic’s defense-in-depth strategy. External partners, including METR and SecureBio, also evaluate the safety of the models.
Source: Read the original article


