Large-scale online deanonymization with LLMs

Researchers demonstrate that LLM agents with internet access can re-identify pseudonymous users (including Hacker News users and Anthropic Interviewer participants) at high precision using only online profiles and conversations — matching hours of human investigator work. Their pipeline extracts identity-relevant features, uses semantic embeddings for candidate matching, then applies LLM reasoning to verify matches at scale. Unlike classical deanonymization (e.g., Netflix Prize), this approach works directly on raw unstructured text across arbitrary platforms.