Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

PrivacyTrace Measures Privacy Leakage in AI Agent Tool Calls

AI summary

Shanghai AI Laboratory, the University of Electronic Science and Technology of China, and the University of Sydney introduced PrivacyTrace, a benchmark that measures how AI agent tool-call traces can expose user attributes. Using 1,000 synthetic users and 67,950 calls across 25 MCP servers, the study reported 26.25% attribute-inference success from a single-domain trace, rising to 32.77% when traces were linked across four domains. TMTPost also notes that request parameters were the largest source of leakage, while the proposed mitigations remain unvalidated.

In-depth insight

The benchmark shows that legitimate tool use can reveal sensitive user profiles even when conversations remain private, making trace retention, parameter handling, and cross-service linkage important AI agent governance concerns.

AI 时代最私密的对话,可能不是你发出的那一句,而是 Agent 帮你发起的那一次工具调用。

你让 AI 助手查下周去伦敦的路线,它顺手调了地图服务;你让它整理体检报告,它连上了健康数据库。你和它的对话内容,理论上只有你和模型厂商能看到。但上海人工智能实验室 AgentDoG 团队最近发布的一项研究提醒我们:即便对话全程保密,AI Agent 在后台调用工具时留下的“作业痕迹”,也足以让观察者拼出你住哪里、信什么教、身体怎么样,勾出一幅相当完整的画像,虽然猜不出你的名字。

这项题为《工具调用能在多大程度上暴露你?来自工具调用轨迹的隐私泄露》的研究,由上海人工智能实验室联合电子科技大学、悉尼大学完成,构建了一个叫 PrivacyTrace 的基准,系统量化了这层此前被忽视的风险。结论并不玄学,而是一组可复现的数字。

#PrivacyTrace#Shanghai AI Laboratory#University of Electronic Science and Technology of China#University of Sydney#DeepSeek-V4-Flash#GPT-5.5#Gemini-3-Flash#MiniMax-M3#Model Context Protocol

Original · 9 min readRead the original
Group chat

No comments yet. Start the conversation.