Anthropic Safeguards Blog – 16 Oct 26 (https://safeguards.anthropic.com/2026/self-assessment/)
Scaling Misuse Detection with Self-Assessment (https://safeguards.anthropic.com/2026/self-assessment/)
Self-assessment lets Claude assess the safety of its own responses. It scales well with model capabilities and helps us identify emergent harms.
Anthropic 会使用对话内容进行封号,实现方法是在用户的对话后面接一个提示,让模型自己判断是不是该封号,跟 Claude 说自己是中国人的基本是确定性的会寄了,就看什么时候