Warning: LLMs Able to De-Anonymize User Accounts on Reddit, Hacker News & Other "Pseudonymous" Platforms; Report Co-Author Expands, Advises

A recent study by AI academics and an Anthropic researcher demonstrates that Large Language Models (LLMs) like Claude and ChatGPT can successfully de-anonymize pseudonymous user accounts on platforms such as Reddit and Hacker News. The attack pipeline extracts identity-relevant features, searches for matches using semantic embeddings, and verifies candidates, achieving 99% precision in connecting Hacker News users to their LinkedIn profiles, and linking different Reddit accounts belonging to the same person based on shared discussions, like favorite movies. Co-author Joshua Swanson noted that any online forum with sufficient publicly searchable text is vulnerable, and preventing misuse is difficult because the techniques rely on common summarization and search tasks. The authors advise users to adopt a stronger privacy mindset, recognizing that the combination of small details across many posts creates a unique fingerprint, and suggest using new accounts for sensitive information.

(Source：Wjamesau Substack)