Large-Scale Online Deanonymization with LLMs
Summary
The article discusses research showing that large language model (LLM) agents can deanonymize anonymous online posts by identifying individuals across platforms with high precision, and that this capability scales to very large candidate pools. It outlines two benchmarking proxies (cross-platform matching and splitting accounts) and notes real-world testing on the Anthropic Interviewer dataset, highlighting significant privacy and security implications along with proposed mitigations for platforms and providers.