Anthropic为Claude引入自我评估机制
Anthropic宣布,Claude在每次回答后会对整段对话生成一次初步安全评估,该技术被称为“自我评估”。Anthropic称自6月起开始使用这一机制,它发现了其他安全措施原本未针对的滥用行为。据中文报道,评估指令与结果不会展示给用户,也不会带入后续对话;结果会被汇总分析,单条评估不会被视为违规定论,被标记的案例会交由更强的分类器或人工复核后才可能采取处置。Anthropic称,在Claude Fable 5上,自我评估标记了更强提示分类器判为潜在有害对话的78%,以及高优先级案例的94%,误报率为1%,并称因Claude已读过对话,该机制执行效率很高。
Anthropic称该机制发现了其他安全措施未针对的滥用行为,并称它依靠Claude自身判断,可随模型能力提升而改进,而非像分类器那样只针对固定类别的危害。单次标记不会导致封号,而是进入汇总复核,使检测链条延伸到对话完成之后。
We’ve trained Claude to flag potential misuse in conversations, a technique we call self-assessment. Right after it responds, Claude gives a preliminary safety assessment of the conversation. Since we began using self-assessment in June, it has surfaced misuse that our other safeguards weren’t designed to catch. Because Claude has already read the conversation, self-assessment is very efficient. And unlike classifiers, which target a discrete set of harms, it draws on Claude’s own judgment, so it gets more accurate as our models improve. Read more on our new Safeguards blog: https://safeguards.anthropic.com/2026/self-assessment/