Yo, if you've been chilling under a throwaway on Reddit, HN or wherever thinking "nah, nobody cares enough to dig" — bad news, bro. ETH Zurich +Anthropic dudes just dropped a bombshell paper showing LLMs straight-up murder online pseudonymity.
They built an agent that grabs your old rants, comments, movie takes — whatever unstructured crap you've posted — pulls out your unique writing style, slang, weird obsessions, then matches it against public profiles(LinkedIn, etc.) or other anon accounts. Cost? $1–$4 per target via API calls.Total experiment ran under 2k bucks.
Key mind-blowers from the paper:
- Open-world agent (full internet access) re-IDs real people from pseudonymous posts with scary precision— replicates hours of human detective work in minutes.
- In closed tests:up to 68% recall at 90% precision linking profiles (vs basically 0% for old-school methods without LLMs).
- Hacker News → LinkedIn:45% recall at 99% precision, 67% at 90%.
- Reddit movie communities cross-matching: harder, but still 8.5% at 90% prec,jumps to 48% if the person yapped about 10+ same movies.
- Temporal split (same Reddit user, posts divided by year): ~33% at 99% prec.
- Even on 33 anonymized Anthropic interview transcripts (censored hard): agent nailed 27% at 82% prec,beating previous best.
- Bonus horror: on Reddit academics under fakes — 52% hit rate.
Bottom line: that "anonymous" shitpost,flame war or dumb joke? It's now a ready-made dossier on you.Any stalker, employer, gov or just bored script-kiddie with API credits can run this at scale.
Pseudonymity is dead as we knew it. Time to rethink how we post online fr.
Paper: https://arxiv.org/pdf/2602.16800
What anon handles are you killing off first? Or am I overreacting? drop thoughts.
If you liked it — follow me, I’d be glad to see you in the feed💪
#AI #Privacy #Deanon #CyberSecurity #LLM #OnlinePrivacy #Anthropic #ETHZurich