Large language models can now unmask pseudonymous users with surprising accuracy, according to new research from computer scientists. The study tested LLM-based deanonymization against traditional methods using Netflix Prize dataset data, finding that LLMs far outperformed classical attacks even when researchers added thousands of decoy identities to confuse the system.

In experiments with 10,000 candidate profiles, LLM approaches maintained high precision while making more guesses, whereas traditional methods failed almost completely at moderate precision levels. The researchers found that even basic LLM search techniques achieved non-trivial recall rates at 99% precision, with more advanced reasoning steps doubling performance. This represents a significant leap in deanonymization capabilities.

The findings raise serious privacy concerns as governments could unmask online critics, corporations could build hyper-targeted advertising profiles, and attackers could launch personalized social engineering scams at scale. The researchers propose several mitigations including API rate limits, automated scraping detection, and guardrails in LLM systems to refuse deanonymization requests.