CoolFace
Datasetpublic

benzeng/reasoning-theft-poc

配套论文:Stealing Reasoning Traces from Proprietary LLM APIs (arXiv:2608.09867,CC BY 4.0,Panfilov et al., 2026)——本仓是该论文的解读与本地最小复现, 不包含论文原文;文中定量结果(315,320 块 / 367 PII / 182 凭证等)均引自论文。 免责声明:本仓为防御性安全研究。论文披露的漏洞在发表前已由相关供应商修复(见论文 Reproducibility Statement);本复现全部在本地小模型(MiniCPM5-2B / Spark-X2.5-1.7B)上进行,使用虚构测试密钥, 未触碰任何真实 API 或真实用户数据。请勿将此类技术用于未授权的系统。 仓库内容:README.md(本报告)· poc_attack.py(复现编排脚本)· serve.py + SERVE.md(带 --encrypt-thinking 加密思维链模式的本地推理服务及手册; 消融 hook 为懒加载,单独此文件即可运行加密模式,配任意 HF 格式本地思维链模型)·… See the full description on the dataset page: https://huggingface.co/datasets/benzeng/reasoning-theft-poc.

sourceHugging Faceapache-2.0updated 14d agoView on Hugging Face
0likes94downloads
3 commits on main
28664d214d ago

加入带 --encrypt-thinking 模式的 serve.py 及使用手册 SERVE.md

benzeng
f0a1d6f14d ago

论文解读+本地复现:加密推理块的跨会话/跨模型解密攻击、消融辅助提取与 session 绑定缓解(arXiv:2608.09867)

benzeng
a310f2114d ago

initial commit

benzeng