LLM Agents Successfully Deanonymize Users from Anonymous Online Posts

New research demonstrates that Large Language Models (LLMs) are highly effective at de-anonymization. The study shows that LLM agents can determine a user's identity from their anonymous online posts with high precision, scaling to tens of thousands of candidates. The method was tested across platforms including Hacker News, Reddit, LinkedIn, and anonymized interview transcripts.

While it has long been understood that individuals can be uniquely identified by surprisingly few attributes, practical application was previously limited. Data often exists in unstructured forms, and deanonymization traditionally required human investigators to manually search and reason based on clues. This research reveals that from just a handful of comments, LLMs can now infer key personal details such as where a user lives, their profession, and their interests. The agents can then use this inferred profile to search for the individual on the web.

The findings indicate that this capability is not only theoretically possible but is becoming increasingly practical, posing a significant new threat to online anonymity. The research highlights a shift from labor-intensive human investigation to automated, scalable AI-driven identification.