Researchers from ETH Zurich, MATS, and Anthropic have published findings demonstrating that large language models can identify individuals behind anonymous online accounts by analyzing their publicly posted content. In tests on Hacker News users who had linked their LinkedIn profiles, the AI system correctly identified 226 out of 338 targets, achieving a 67% success rate.
The attack method, titled "Large-scale online deanonymization with LLMs," relies on four steps: extracting biographical details from posts, converting that information into searchable data embeddings, using a language model to reason through candidate matches, and calibrating confidence levels. The entire process costs between $1 and $4 per target and requires no hacking, data breaches, or insider access.
How the technique works:
- Extract: AI summarizes anonymous posts to identify details like location, employment, hobbies, and writing patterns
- Search: The summary is converted to an embedding to find similar profiles among thousands of candidates
- Reason: A more powerful model cross-checks the best matches against extracted details
- Calibrate: The system rates its own confidence and avoids guessing when uncertain
The researchers noted limitations in their study. Their test subjects were accounts that had already revealed their real identities through linked profiles or deliberately split post histories, representing controlled conditions rather than proof that arbitrary pseudonymous accounts can be reliably cracked. When tested against a pool of 89,000 candidates, the strongest method achieved approximately 50% accuracy at 90% precision.
The authors withheld their code, prompts, and all real identities discovered during the research. The study passed through ETH Zurich's ethics review board before publication.
The findings add to existing concerns about online privacy. Users who post pseudonymously for activism, privacy protection, or professional reasons may face heightened risk, as scattered identifying details across years of comments can collectively form a unique fingerprint. The study appears several months after disclosure that state-backed hackers used Claude for cyberespionage operations.
