1
LLMs and Contextual Integrity
大模型持久记忆暗藏隐私雷区，用情境完整性框架守住信息边界，AI安全必修课。
I have been thinking a lot about AI and integrity. Part of that is contextual integrity. I recently found two papers on the topic. “ CIMemories:…
2
ChatGPT’s Computer History tracks your clicks and keystrokes
ChatGPT的Computer History默默记录你的点击和键盘，隐私边界再受考验，AI助手贴心还是越界？
ChatGPT's desktop app on macOS has a new feature called Computer History that turns your actions into training data, learning how you work, suggesting…
3
Ask HN: Does AI watermarking present a new attack vector?
AI水印可能藏着用户隐私泄露与攻击风险，写代码时是否正在留下可追踪指纹？值得深思。
One thing I have not been able to determine from Claude's documentation on watermarking[0] is what kind of metadata they store in association with a g…
4
Can Released LLM Vocabularies Support Token-Level Estimation of Hidden Corpora?
探究大模型公开词表能否反推隐藏训练语料的token分布，直击数据隐私与安全痛点
arXiv:2608.10690v1 Announce Type: new Abstract: Pretraining corpus composition shapes LLM capabilities, but it often remains hidden even when model we…
5
Stealing Reasoning Traces from Proprietary LLM APIs
一条加密token泄露链，暴露闭源大模型推理过程可被逆向窃取，安全边界再受拷问。
Stealing Reasoning Traces from Proprietary LLM APIs A vanity domain name ( stolen-thoughts.com ) for a neat paper : Anthropic, OpenAI, and Google retu…
6
AI coding assistant is quietly shipping your secrets
AI编程助手正悄悄把你的密钥发到公网，33个文件、30个包里藏着你的API Key——用AI写代码前必看的安全警示。
Article URL: https://reykur.io/blog/ai-coding-assistant-shipping-secrets/ Comments URL: https://news.ycombinator.com/item?id=48815447 Points: 1 # Comm…
7
Claude Code is steganographically marking requests
开发者自查发现 Claude Code 竟在请求中隐写标记，隐私风险值得每个 AI 工具用户警惕。
Article URL: https://thereallo.dev/blog/claude-code-prompt-steganography Comments URL: https://news.ycombinator.com/item?id=48734373 Points: 783 # Com…
8
Disparate privacy risks from medical AI
医疗AI在提升诊断可及性的同时，训练数据暗藏患者隐私泄露风险，Nature揭示攻击面与防护挑战
Nature, Published online: 24 June 2026; doi:10.1038/s41586-026-10688-0 AI models for medical diagnostics are vulnerable to membership inference attack…
9
Exposing the Illusion of Erasure in Knowledge Editing for LLMs
揭露LLM知识编辑的“擦除幻觉”，26页报告揭示技术漏洞，拆解AI修改知识的真实困境。
arXiv:2606.23276v1 Announce Type: cross Abstract: Knowledge Editing (KE) has emerged as a frontier for updating specific facts in LLMs without costly …
10
An Evaluation of Data Leakage Risks in Tool-Using LLM Agents in Realistic Scenarios
首份系统评估LLM智能体在工具使用场景下的数据泄露风险，揭示API调用和记忆机制中的严重隐私漏洞。
arXiv:2606.17114v1 Announce Type: cross Abstract: AI agents are increasingly being adopted in enterprise and personal settings with access to emails, …
11
If you build apps with AI tools, this is .uh concerning
用AI工具构建应用时，隐藏的安全隐患值得警惕
Article URL: https://substack.com/profile/173863161-dan-cochran/note/c-274040959 Comments URL: https://news.ycombinator.com/item?id=48481349 Points: 2…
12
Cloudflare Turnstile requiring fingerprintable WebGL
Cloudflare Turnstile用WebGL指纹实现无感验证，但隐私风险需警惕。
Article URL: https://hacktivis.me/articles/cloudflare-turnstile-webgl-fingerprinting Comments URL: https://news.ycombinator.com/item?id=48345840 Point…
13
When Helpful Context Leaks: Privacy Risks in Domain-Adapted ASR
领域自适应ASR的上下文竟成隐私漏洞？新研究揭示“帮助”背后的泄露风险。
arXiv:2605.28211v1 Announce Type: new Abstract: SpeechLLMs are increasingly deployed in professional settings where domain customisation is standard p…