Skip to content

Translation in progress Some entries still include text in the original language.

Embrace The Red·Indexed by this site · Originally published

博主复现加密 LLM 推理轨迹还原攻击,跨账号取回 GPT-5.6 推理中的密码

Recovering Encrypted LLM Reasoning Traces

AI summary

安全研究者 Johann 复现了论文《Stealing Reasoning Traces from Proprietary LLM APIs》提出的攻击,把加密推理块回放到同一厂商较易越狱的模型并诱导其转述,成功还原 OpenAI 推理轨迹中的语义内容。

Related papers
Read originalembracethered.com