Large-Scale Online Deanonymization with LLMs
LLMdeanonymizationprivacysecurityartificial intelligencemachine learningonline anonymityresearchdata privacyidentity disclosure
A research paper demonstrates that LLM agents can identify individuals from anonymous online posts across platforms like Hacker News, Reddit, LinkedIn, and anonymized interview transcripts with high precision. The method scales to tens of thousands of candidates and infers personal details such as location, profession, and interests from a small number of comments. The study shows that LLMs can automate deanonymization by searching and reasoning based on unstructured data. The research was conducted by MATS Research, ETH Zürich, and Anthropic.