LLM-Assisted Deanonymization

This post summarizes research demonstrating that large language models can deanonymize users from a handful of anonymous online posts across platforms (Hacker News, Reddit, LinkedIn, and interview transcripts). The authors show LLM agents can infer location, occupation, and interests and then search to identify individuals at scale, highlighting a growing privacy and re-identification risk rather than reporting an active cyber incident.