模型与实验室 4.0 · 优秀 2026-05-29 · 文章

Fooling around with encrypted reasoning blobs

Matthew Green 在 2608.09867 论文公开前两个月写下的背景笔记,把 encrypted reasoning blob 视作把服务端保密负担转嫁给客户端的工程决策,并指出 blob 在客户端持有期间能被复制再分发跨请求复用这是后续跨 session 可移植攻击和 decryption 漏洞的前提该笔记被 8-11 的 kotekjedi thread 与论文引用,作为密码学界对 frontier 模型架构选择的早期标注

打开原文回到归档

Fooling around with encrypted reasoning blobs

中文摘要

Matthew Green 在 2608.09867 论文公开前两个月写下的背景笔记,把 encrypted reasoning blob 视作把服务端保密负担转嫁给客户端的工程决策,并指出 blob 在客户端持有期间能被复制再分发跨请求复用这是后续跨 session 可移植攻击和 decryption 漏洞的前提该笔记被 8-11 的 kotekjedi thread 与论文引用,作为密码学界对 frontier 模型架构选择的早期标注

English Summary

Matthew Green's May 2026 background post predates the 2608.09867 paper by two months. He frames the encrypted-reasoning-blob design as shifting the confidentiality contract from the server to the client, and flags that blobs held by the client can be copied, redistributed, and reused across requests - the precondition the later paper turns into a portable cross-session attack and decryption jailbreak. The post is the cryptographic community's earlier marker of the same architectural trade-off.

来源 / Obsidian 引用

  • 本机 Obsidian 同日 digest 落地(ClawFeed 24h / AK-RSS-Digest 89 源 / X-Hot-Brief / 内容选题编排)
  • 评分依据:原文正文(不靠标题/摘要/源声誉)