Pseudonymity's subscription fee just dropped to $2 🤖

Researchers from ETH Zurich, the AI safety group MATS, and Anthropic have published a study showing that large language models can identify the real-world owners of pseudonymous online accounts for between $1 and $4 in compute costs per search. The paper, titled "Large-scale online deanonymization with LLMs," was originally released in February and resurfaced on X and Reddit on September 23, 2026.

The team built a fully autonomous pipeline that reads a target's posts and attempts to match them to a real identity. It operates in four stages: "Extract," which summarizes details such as location, occupation, hobbies, or linguistic quirks; "Search," which converts that summary into a numerical embedding to scan candidate profiles; "Reason," in which a more capable model cross-checks the top candidates; and "Calibrate," which forces the system to abstain unless it crosses a confidence threshold.

In a test on 338 Hacker News users who had linked a LinkedIn profile in their bio, the system correctly named 226, roughly 67%, and was wrong on about one in ten of the guesses it actually emitted. In a separate run on transcripts from an Anthropic interview study with 125 scientists, the model correctly identified at least 9 individuals from how they described their work.

Co-author Nicholas Carlini is a researcher at Anthropic, the company behind the Claude model. The paper's quoted conclusion reads: "Our results show that the practical obscurity protecting pseudonymous users online no longer holds and that threat models for online privacy need to be reconsidered." The team stated that the attack does not require a data breach, relying only on web search and reasoning capabilities already available in commercial chatbots.