Researchers Demonstrate LLMs Deanonymize Online Accounts
Researchers from ETH Zurich, Anthropic and the Machine Learning Alignment and Theory Scholars program publish a non-peer-reviewed study showing LLM-based AI agents can deanonymize online accounts. In experiments on Hacker News, LinkedIn, Anthropic interviews and split Reddit accounts the system identified up to 68% of matching accounts at 90% precision and cost under $2,000. The authors warn this automation increases reidentification risks for pseudonymous users.
Key Points
- 1Demonstrate LLM-based agents identify anonymized accounts up to 68% recall at 90% precision
- 2Show automation vastly outperforms non-LLM methods, enabling large-scale, fast deanonymization
- 3Warn that pseudonymous users, journalists, and activists face increased reidentification risk; adopt stricter practices
Scoring Rationale
Strong, novel demonstration of scalable LLM deanonymization; limited by non-peer-reviewed methods and lab-constrained, curated datasets.
Sources
Public references used for this report.
Practice with real Logistics & Shipping data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Logistics & Shipping problems
