LLMs Perform Large-Scale Online Deanonymization Against Pseudonymous Users

Researchers led by Simon Lermen and collaborators from MATS Research, ETH Zurich, and Anthropic publish a pre-press paper demonstrating that large language models can deanonymize internet users by linking pseudonymous posts to real profiles across Hacker News, Reddit, LinkedIn, and interview transcripts. In experiments the method identified 226 of 338 Hacker News targets (67% recall) at 90% precision, costing about $2,000 total; the authors warn this enables scalable, affordable deanonymization.