ETH Zurich and Anthropic built an AI that doxes anonymous accounts in a minute for $2
skdh · x · 2026-09-25
Researchers from ETH Zurich and Anthropic published a paper showing that practical anonymity online is effectively dead. They built a fully autonomous AI pipeline that takes pseudonymous posts, extracts identity signals, and searches the web to pinpoint a user's real identity in under a minute, at about $2 per target, with no human investigator involved.
Key findings:
- 67% of Hacker News users were correctly identified
- When the system makes a guess, it's right 90% of the time
- It even unmasked scientists whose interview transcripts had been explicitly redacted for privacy
- Users who took a full year off and changed their interests were still re-linked to old profiles with 90% precision
The takeaway: persona changes and time gaps no longer protect pseudonymity, a serious warning for privacy and security.
More from Safety
- X's report system can't catch ML-community scammers who impersonate people, not accounts — gandamu_ml · 2026-09-25
- China and the US race ahead on AI — can they manage the risks together? — scmp_news · 2026-09-25
- Quillette essay argues AI safety advocates are dangerous for dismissing technical security — basedjensen · 2026-09-25
- CVE-2025-13032: Researchers Detail Breaking Into the Avast Antivirus Sandbox, Part 2 — jedisct1 · 2026-09-25
- YouTube extends likeness detection to voices as AI infrastructure spend hits $10.3T by 2032 — emmanuelvivier · 2026-09-25
- OpenAI reportedly prepping GPT-6 Cyber security model for DevDay — emmanuelvivier · 2026-09-25