In partnership with
Before we start: If you sell anything, check this out:
Stop losing deals in between meetings.
Aligned is changing how B2B teams run their most important deals.
Most reps guess where a deal really stands. They forecast on a gut feeling and the buyer’s word vs what's actually happening. Then they realize the deal slipped through only after it's too late. Aligned surfaces what buyers are actually doing inside every deal: the risks, the openings, and the exact next step to keep things moving.
Start increasing your close rates and stand up your first deal room in minutes.
Let’s go.
Your secret account isn't that secret
You probably have at least one account with no name on it. An old Reddit, or a forum handle nobody at work knows about.
A paper from February says that account isn't as anonymous as you think. It comes from researchers at ETH Zurich plus Nicholas Carlini at Anthropic, who's spent years finding ways to break AI models.
Most people missed it. Which is a shame, because it's about you and the stuff you post online that you don’t want other people to know about.
The short version: an AI can now read what you post anonymously and figure out who you are, for a few dollars per person.
Finding you used to be too much work
Unmasking someone online was always possible. It just took hours of reading and cross-checking per person, so unless you were famous or in trouble, nobody bothered.
The famous exception was the Netflix Prize attack, which matched "anonymous" movie ratings to real IMDb users. That worked because ratings are tidy data: a movie, a score, a date.
Your posts are messy. Until now, that mess was your protection.
LLMs read mess just fine. The system in this paper pulls clues out of your writing, searches a huge pool of profiles for people who fit, then checks the best candidates.
Say you've mentioned working in logistics and playing in a Sunday league. Neither gives you away alone. Next to a LinkedIn page that says the same things, it's a fingerprint.
Given web access and Anthropic Interviewer transcripts published with names removed, it could sometimes name the actual scientist from a single interview.
How it stacks up against a human investigator
A good human could do everything this agent does. The difference is scale.
The agent named 226 of 338 Hacker News users. When it made a guess, it was right about 9 times out of 10 (!!), and when it wasn't sure it mostly said so instead of guessing.
Each profile cost the team between $1 and $4, running off-the-shelf models like GPT-5.2 and Gemini 3.
These were test setups where the researchers already knew the right answers. And results vary a lot: matching people across Reddit movie communities was much harder, around 8.5% at the same accuracy, per the full paper.
So "anonymity is dead" is too strong. The accurate version: unmasking someone used to cost an afternoon, and now it costs less than a coffee.
When you should actually care
Say you've had a Reddit account since college. You never posted your name. But over the years you mentioned your city, your job, your gym and that trip to Lisbon.
Your LinkedIn and Instagram say a lot of the same things.
That overlap is exactly what this attack feeds on. No single comment gives you away. The pattern does.
Care if you post anonymously about your employer or your health, or anything you'd hate to see tied to your name. A tool like Redact can wipe or scramble your old Reddit comments, and the free tier covers the basics.
For the bigger picture, the EFF's Surveillance Self-Defense guide is a good free read.
You can relax if everything you post anonymously is stuff you'd happily say under your own name.
Before you panic
The researchers built their own test sets and graded their own results. To their credit, they didn't unmask real anonymous users.
They took profiles with known owners, stripped the names and tested on those. They also kept their code and prompts private.
And it didn't come out of nowhere. A January paper had already re-identified people in the Anthropic Interviewer data.
Since then others have tested similar attacks and started building defenses against them.
Arvind Narayanan, one of the Netflix Prize researchers, has been warning about this for years. The difference now is the price tag.
Yes, this one was a bit technical with papers and research. But that’s where some of the most important AI stories are hiding. We find them and dissect them for you.
See you tomorrow.
Stefano Martell
Co-Founder, The AI Leverage by Bestapps.ai
PS - Latest Harvard Report on AI Customer Service:
New Harvard Business Review Analytic Services report: why most AI tools fail at B2B customer service, and what to ask before you invest. Get the free report.
