Anthropic says it trained Claude to flag possible misuse in its own conversations

Anthropic described a safeguard it calls self-assessment: right after Claude responds, it gives a preliminary safety assessment of the conversation, flagging potential misuse. Everything below is Anthropic's own account, from its post and a new Safeguards blog.

• Since June: Anthropic says it has used self-assessment since June and that it has surfaced misuse its other safeguards were not designed to catch.
• Why it says it works: Claude has already read the conversation, so the check is efficient. Unlike classifiers, which target a fixed set of harms, it draws on Claude's own judgment, so Anthropic says it becomes more accurate as its models improve.