📄 中文摘要
2026年2月，研究者西蒙·勒尔曼的研究表明，具有互联网访问能力的大型语言模型（LLM）能够通过交叉引用写作风格、行为模式和公开数据，以超过85%的准确率去匿名化在线用户。这一能力的转变不仅是理论上的漏洞，更是对“匿名”在网络上实际含义的深刻影响。LLM去匿名化已成为一种生产规模的威胁，而非仅仅是研究中的好奇现象，任何具备互联网访问和足够上下文的部署都可能面临这一风险。
大型语言模型去匿名化正在暴露在线真实身份
出处: LLM Deanonymization Is Exposing Real Identities Online
发布: 2026年3月2日
📄 English Summary
LLM Deanonymization Is Exposing Real Identities Online
In February 2026, a study by Simon Lermen demonstrated that large language models (LLMs) with internet access can reliably deanonymize online users with over 85% accuracy by cross-referencing writing styles, behavioral patterns, and publicly available data. This capability shift is not merely a theoretical vulnerability; it fundamentally alters the meaning of 'anonymity' online. LLM deanonymization has emerged as a production-scale threat rather than a mere research curiosity, posing risks to any deployment with internet access and sufficient contextual information.