Translation in progress Some entries still include text in the original language.
Simon Willison·Indexed by this site · Originally publishedSelectedAttention53
论文披露从专有 LLM API 窃取加密思维链的方法
Stealing Reasoning Traces from Proprietary LLM APIs
AI summary
一篇论文发现 Anthropic、OpenAI 和 Google 返回给客户端的加密思维链块可在不同会话、用户和模型间重放,把前沿模型的推理轨迹喂给同族较弱模型并越狱后,即可还原出强模型的隐藏推理明文。
Why it was selected
论文披露加密思维链块可在同族模型间重放并越狱还原,还衍生出借思维链实施提示注入的新变体。
Read originalsimonwillison.net