Key facts
- An AI pipeline can deanonymize online accounts by matching posts to real-world identities.
- The AI uses web search, embeddings, and reasoning models like GPT-5.2.
- The cost per target is estimated to be between $1 and $4.
- The process does not involve hacking or data breaches.
- Researchers achieved 67% accuracy in identifying users from anonymized Hacker News accounts linked to LinkedIn.
- The study authors withheld their code, prompts, and uncovered identities.
Researchers from ETH Zurich, MATS, and Anthropic have developed an AI pipeline that can deanonymize online accounts with high precision, potentially signaling the end of practical internet anonymity. The system, detailed in a study titled "Large-scale online deanonymization with LLMs," uses publicly available information and reasoning models to match pseudonymous posts to real-world identities.
The AI pipeline operates in four steps: Extract, Search, Reason, and Calibrate. First, it extracts identifying details from a user's posts, such as location, job, hobbies, or even linguistic quirks. This summary is then converted into an embedding for searching similar profiles. A powerful reasoning model cross-checks potential matches, and a calibration step ensures the AI only makes a guess when it has high confidence.
In one test, the AI correctly identified 67% of users from anonymized Hacker News accounts that had linked to their LinkedIn profiles, with a precision of 90%. A separate test on interview transcripts from Anthropic correctly identified at least nine scientists based on their work descriptions. The cost for this process is estimated to be between $1 and $4 per target, and it requires no hacking or data breaches, relying instead on the existing capabilities of large language models.
The study's authors have withheld their code, prompts, and any real identities uncovered, and the research underwent an ethics review. They argue that the findings highlight a significant shift in online privacy, as the practical obscurity previously protecting pseudonymous users is no longer sufficient. The research comes after Anthropic disclosed a cyberespionage campaign that used its Claude AI, underscoring concerns about AI misuse.