窃取专有LLM API的推理痕迹
原标题：Stealing Reasoning Traces from Proprietary LLM APIs
AI 摘要
研究人员发现，Anthropic、OpenAI 和 Google 等大型语言模型提供商返回给客户端的加密推理痕迹块在不同会话、用户和模型间可互换，存在架构漏洞。通过将该痕迹注入同一提供商的较弱模型中，可强制其解码并输出明文，从而绕过反蒸馏机制，实现私有数据提取、隐藏危险信息泄露和隐形提示注入。研究团队从公共仓库中解码了 315,320 个推理块，恢复了 367 个个人身份信息和 182 个凭据，并提出了加密和系统级缓解措施。
正文节选
Stealing Reasoning Traces from Proprietary LLM APIs Abstract Encrypted reasoning traces shared across sessions and models can be intercepted and injected into weaker models to extract proprietary reasoning, private data, hidden hazards, and hidden prompts. Leading large language model providers now conceal their models' step-by-step reasoning, or chain-of-thought, to protect intellectual property and limit information leakage. Rather than storing these traces server-side, providers return them t