LLMs Enable Large-Scale Deanonymisation Of Users

Researchers at ETH Zurich, the ML Alignment Theory Scholars program, and Anthropic posted an arXiv preprint describing a pipeline that uses commercial LLM APIs to deanonymise pseudonymous online accounts for as little as $1.41 per target. The ESRC (Extract, Search, Reason, Calibrate) method achieved 45.1% recall at 99% precision linking Hacker News to LinkedIn across 89,000 users and projects roughly 35% recall at 90% precision at million-scale, highlighting platform privacy risks.