Cases › leaked private data › LR-2026-0061
LLM pipeline deanonymizes pseudonymous Hacker News and scientist accounts
Whose AI: GPT-5.2leaked private data
Severity moderate: Other parties' data or systems were touched, and a source states how far it reached. The study deanonymized 338 Hacker News users, a bounded scope of third parties.
Earliest source we hold 2026-09-24
Researchers from ETH Zurich, MATS, and Anthropic built an AI pipeline that matched anonymized Hacker News accounts to real LinkedIn profiles at 90% precision.
Risk grouping: AI used to exploit
| Who caused it | A person or organisation “Researchers from ETH Zurich, MATS, and Anthropic built an AI pipeline that matched anonymized Hacker News accounts to real LinkedIn profiles at 90% precision.” source |
|---|---|
| Intended or not | Intended “In one test, the team gathered 338 users of the mostly pseudonymous Hacker News message board who had linked their LinkedIn profile in their bio, stripped out anything identifying, then asked an AI agent to figure out who they were using nothing but web search.” source |
| Before or after release | Before release “The researchers also didn't publish their tools, prompts, or any of the real names they uncovered, and the study went through an ethics review board before release.” source |
| The AI's part | Used as a tool “In one test, the team gathered 338 users of the mostly pseudonymous Hacker News message board who had linked their LinkedIn profile in their bio, stripped out anything identifying, then asked an AI agent to figure out who they were using nothing but web search.” source |
The story
1 · What it was doingResearchers at ETH Zurich, MATS and Anthropic built a fully autonomous pipeline using web search, embeddings and reasoning models such as GPT-5.2 to identify anonymous Hacker News users from their public posts.
2 · What went wrongThe pipeline deanonymized pseudonymous users, breaching the practical obscurity that protects anonymous accounts online.
3 · What it reached or didIt correctly named 226 of 338 Hacker News users and at least 9 of 125 scientists from Anthropic interview transcripts, exposing real identities.
4 · What was doneThe researchers withheld their code, prompts and every real identity uncovered, and the study passed ETH Zurich's ethics review before publication.
Show every field, with its source and grade
The AI system
Model
GPT-5.2
C one secondary source states it
Tools it was given
web search
C one secondary source states it
How much it acted alone
it acted alone, unwatched
D inferred by us from the described behaviour, reason shown
Stage of use
in research
D inferred by us from the described behaviour, reason shown
Stage of the model
in testing
D inferred by us from the described behaviour, reason shown
How it was connected
agentic with tools
D inferred by us from the described behaviour, reason shown
What it was for
to figure out who anonymous users were using nothing but web search
C one secondary source states it
Meant for live use
no, it was not meant for live use
D inferred by us from the described behaviour, reason shown
What happened
Earliest source we hold
2026-09-24
A the date the source itself carries
Themes
deanonymization, privacy, LLM, web search
D inferred by us from the described behaviour, reason shown
Who caused it
the AI system itself
D inferred by us from the described behaviour, reason shown
Where its input came from
public posts from anonymous accounts
C one secondary source states it
What was done
The researchers withheld their code, prompts, and every real identity they uncovered.
C one secondary source states it
Earlier warning signs
2008 Netflix dataset deanonymization; 2025 Coinbase data breach
C one secondary source states it
What went wrong
leaked private data
D inferred by us from the described behaviour, reason shown
AI agent
an AI agent acted
C one secondary source states it
What it did
The pipeline deanonymized users by matching public posts to real identities.
D inferred by us from the described behaviour, reason shown
Impact
Severity
moderate
D inferred by us from the described behaviour, reason shown
Kind of harm
privacy
D inferred by us from the described behaviour, reason shown
Who was affected
338 users of the mostly pseudonymous Hacker News message board
C one secondary source states it
Kinds of data exposed
personal information
D inferred by us from the described behaviour, reason shown
Why this severity
Other parties' data or systems were touched, and a source states how far it reached. The study deanonymized 338 Hacker News users, a bounded scope of third parties.
D inferred by us from the described behaviour, reason shown
Context
Investigated
by an evaluator working inside
C one secondary source states it
Redacted
code, prompts and real identities withheld
C one secondary source states it
LR-2026-0061 · version 6 · published 2026-10-25 after automatic checks; not yet checked field by field by a person · changes to this case · JSON