February 25, 2026
Whoops—your alt got unmasked
Large-scale online deanonymization with LLM agents
AI says it can unmask your anonymous posts — the internet screams “duh” and “yikes”
TLDR: Researchers say AI can link “anonymous” posts to real identities by inferring details and cross-searching platforms, showing it scales beyond one-off snooping. Commenters split between “this was obvious,” privacy panic, gallows humor, and DIY defenses like rewriting style — warning that AI-powered people search is here now.
Internet drama alert: researchers claim AI “agents” can spot who you are from a handful of “anonymous” posts by guessing your city, job, and hobbies — then hunting you down across platforms. Their tests paired anonymized Hacker News accounts to LinkedIn profiles and even linked split Reddit histories, saying it works best when AI both searches and “reasons.” They also tried a real-world check on anonymized interview transcripts. Translation: what used to take a nosy human now scales to thousands.
Cue the comments. The top mood? A snarky split between “this is obvious” and “we’re doomed.” One skeptic deadpans that matching a Hacker News account that literally links to LinkedIn is… groundbreaking. Others go full Black Mirror, warning this speeds up creepy “people search.” A privacy-minded commenter floats a counter-move: rewrite your posts with a local AI to scramble your writing style — a nod to old-school stylometry, the science of identifying people by how they write.
There’s meta-humor too: “Hello, LLM!” jokes one user, noting that Reddit and HN likely fed the very models now unmasking them. Another quips, “Good thing I always lie on the internet,” while a different commenter claims Microsoft’s Copilot can already summarize a user’s vibe, not just their posts. The researchers argue that exposing the risk helps platforms and labs fix it, but readers are split: shine a light or teach the stalkers? Either way, the internet’s new rule is trending: post like your boss — and your future self — is watching.
Key Points
- •Researchers demonstrate that LLM agents can deanonymize users by inferring attributes from a few posts and searching the web to identify them.
- •A cross-platform proxy benchmark matches anonymized Hacker News accounts to LinkedIn identities using embeddings retrieval plus LLM reasoning, achieving high precision.
- •A split-account proxy on Reddit (temporal or by subreddit) shows LLM embeddings with reasoning outperform Netflix-Prize-style baselines based on activity patterns and metadata.
- •A real-world deanonymization attack on the Anthropic Interviewer dataset was conducted with manual verification to validate effectiveness.
- •The work highlights privacy risks of scalable AI “people search,” discusses potential protections for individuals and responses by platforms and AI labs, and justifies publication despite misuse risks.