# Large-scale online deanonymization with LLMs

**Authors:** Simon Lermen, Daniel Paleka, Joshua Swanson, Michael Aerni, Nicholas Carlini, Florian Tramèr  
**arXiv:** [2602.16800](https://arxiv.org/abs/2602.16800)  
**Date:** 2026-02-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2602.16800v1/

## Key points

- Researchers explore using large language models (LLMs) for at-scale deanonymization
- Agent accurately re-identifies Hacker News users and Anthropic Interviewer participants using only pseudonymous online profiles and conversations
- Attacks developed for closed-world setting utilizing databases of pseudonymous individuals with unstructured text
- Scalable attack pipeline leverages LLMs to extract identity-relevant features, search for candidate matches through semantic embeddings, and reason over top candidates to verify matches and reduce false positives
- Approach operates directly on raw user content across various platforms without structured data or manual feature engineering
- Evaluation conducted on three datasets with known ground-truth data, showing LLM-based methods significantly outperform classical baselines with recall rates up to 68% at 90% precision
- Results highlight the ineffectiveness of practical obscurity protecting pseudonymous users online and emphasize the need to reassess threat models concerning online privacy

## Summary

In their research project titled "Large-scale online deanonymization with LLMs," Simon Lermen, Daniel Paleka, Joshua Swanson, Michael Aerni, Nicholas Carlini, and Florian Tramèr explore the use of large language models (LLMs) for performing at-scale deanonymization. They demonstrate that with full Internet access, their agent can accurately re-identify Hacker News users and Anthropic Interviewer participants using only pseudonymous online profiles and conversations. This task would typically require hours for a human investigator. The researchers further develop attacks for the closed-world setting by utilizing two databases of pseudonymous individuals containing unstructured text related to each individual. They implement a scalable attack pipeline leveraging LLMs to extract identity-relevant features, search for candidate matches through semantic embeddings, and reason over top candidates to verify matches and reduce false positives. This approach stands out from prior deanonymization work as it operates directly on raw user content across various platforms without the need for structured data or manual feature engineering. To evaluate their attacks, three datasets with known ground-truth data are constructed. The first dataset links Hacker News users to their LinkedIn profiles based on cross-platform references present in the profiles. The second dataset involves matching users across Reddit movie discussion communities while the third dataset splits a single user's Reddit history in time to create two pseudonymous profiles for matching purposes. In each scenario, the LLM-based methods significantly outperform classical baselines achieving recall rates of up to 68% at 90% precision compared to near 0% for non-LLM methods. These results highlight that the practical obscurity protecting pseudonymous users online is no longer effective and emphasize the need to reassess threat models concerning online privacy. By showcasing the capabilities of LLMs in deanonymization tasks across diverse platforms and datasets, this research underscores the importance of addressing privacy concerns in an increasingly interconnected digital landscape.