The Firm Screening OpenAI's Users Was Filing Reports to the Feds

A researcher found 53MB of Persona's government codebase exposed online. What's inside is bigger than age verification.

State of Surveillance Editorial Team

Digital face recognition scan with green data overlay
Photo via Unsplash

TL;DR:

  • 53 megabytes of Persona's government frontend was found sitting exposed on a public, FedRAMP government endpoint. No hacking required: the full source code was readable through leftover developer source maps.[1][2]
  • The biggest client implicated is OpenAI. Persona runs ChatGPT's identity verification and, by its own marketing, "automatically screens over 99% of users behind the scenes." That screening ran for roughly 18 months before OpenAI publicly told anyone about ID checks.[2]
  • 269 verification checks across 14 check types are defined in the code: facial recognition against public figures, "suspicious entity" flagging, deceased-record matching, sanctions screening, and more. These are checks that exist in the system, not 269 things run on every single person.[1][4]
  • It can file reports straight to financial-crime agencies: the code contains modules that file Suspicious Activity Reports (SARs) to FinCEN in the US and Suspicious Transaction Reports (STRs) to FINTRAC in Canada.[1]
  • Peter Thiel connection: Persona is backed by Founders Fund, Thiel's venture firm, the same money behind Palantir.[1][3]
  • An unproven infrastructure echo: the leaked platform sits on an "ONYX" subdomain, which happens to share a name with an ICE surveillance tool. The code itself contains zero references to ICE or immigration enforcement. Treat this as a coincidence-shaped question mark, not a proven pipeline.[1]

What Got Exposed

On February 19, 2026, a security researcher who goes by Celeste (handle vmfunc) published a report that should make anyone who has ever uploaded an ID for "age verification" sit up.[1]

Persona's entire government dashboard frontend, 53 megabytes across 2,456 TypeScript source files, was readable on a public endpoint. No exploit. No breach. Just leftover developer source maps at a /vite-dev/ path on app.onyx.withpersona-gov.com, an unauthenticated, FedRAMP-authorized government endpoint. The kind of thing that is supposed to be the most locked-down infrastructure a company runs.[1][2]

Source maps are the developer's-eye view of compiled code: variable names, file structure, the works. Leaving them on production is like shipping a building with the architectural blueprints taped to the front door. What was inside the blueprints is the actual story.

Who Is Persona?

Persona is a San Francisco identity verification company. When a platform needs to confirm you're a real human, the right age, or not on some list, a lot of those checks run through Persona behind the scenes.

Its backers include Founders Fund, Peter Thiel's venture capital firm: the same Thiel who co-founded Palantir. Founders Fund led Persona's $150 million and $200 million funding rounds.[1][3]

Persona sells itself as a simple identity check. The exposed code describes something much larger.

The OpenAI Angle

Here's the part that got buried under the age-verification framing: the largest client implicated in the leak is OpenAI.[2]

Persona handles identity verification for OpenAI's products. By Persona's own marketing language, it "screens millions monthly" and "automatically screens over 99% of users behind the scenes." In plain terms: if you use ChatGPT in a context that triggers verification, you're likely passing through this machine without ever clicking a button that says so.[2]

The timeline is the uncomfortable part. The watchlist-style screening service appears to have been running for roughly 18 months before OpenAI publicly disclosed that it does ID checks at all. Users were being run through identity screening long before they were told it existed.[2]

269 Checks, 14 Types

The code defines a CheckName enum: a master list of 269 individual verification checks, grouped into 14 check types. Important nuance, because the original framing of this story got it wrong: these are checks that exist in the system, not 269 things that run on every user. Think of it as a menu, not a meal everyone is force-fed.[1][4]

Still, the menu is the point. Among the checks named in the code:

  • Facial recognition against public figures: a public-figure detection check that compares a selfie against known faces[1]
  • SelfieSuspiciousEntityDetection: a check that flags a face as "suspicious" by criteria the code does not explain to the user[1]
  • Deceased-record matching: cross-referencing against Social Security Administration death records[1]
  • Sanctions and watchlist screening: checks against sanctions-style lists[1][4]
  • Adverse-media style screening: checks tied to negative news and risk signals[4]
  • Document and age consistency: comparing claimed details against the ID and the face[1]

This is a biometric and watchlist screening stack sitting behind a product most people think of as "prove you're 18."[4]

A Direct Line to the Feds

The detail most users would never guess: the code includes modules that file reports straight to government financial-crime agencies.[1]

Specifically, it can file Suspicious Activity Reports (SARs) to FinCEN, the US Treasury's Financial Crimes Enforcement Network, and Suspicious Transaction Reports (STRs) to FINTRAC in Canada. These are anti-money-laundering and KYC mechanisms: the same kind of report a bank files when it suspects something off about a transaction. The point isn't that an identity vendor handles money-laundering signals; it's that the thing branded to you as "verify your age" is wired into the federal financial-crime reporting system.[1]

The government version of Persona's platform also includes capabilities its commercial offering does not advertise: real-time monitoring, behavior tracking on government identity systems, and the SAR/STR filing pipeline running on government infrastructure that handles PII and biometrics.[1]

The ONYX Question (Read This Carefully)

You will see this story framed as "Persona feeds ICE." Be skeptical of that, including when we'd love it to be true.

The honest version: the leaked government platform lives on an "ONYX" subdomain, and "ONYX" is also the name of a Fivecast surveillance tool associated with ICE. That naming echo is the entire basis for the ICE connection. The leaked code itself contains zero direct references to Fivecast, ICE, immigration enforcement, or deportation.[1]

So this is an infrastructure correlation, not a proven data pipeline. It's a question worth asking out loud. It is not a fact, and we're not going to pretend it is one.

Who Sends Your Data to Persona

Persona's reach is wide. Beyond OpenAI, Persona-style age and identity verification shows up across major consumer platforms, including ones in the Discord/Reddit/Roblox tier of "you've definitely used at least one of these." When a platform asks you to verify your age or identity, your selfie and ID often flow into infrastructure like this.[2]

One thing to be precise about: an earlier version of this story claimed Discord publicly announced it "will not be proceeding with Persona." That claim is not supported by the primary reporting, so we've pulled it. Discord and other platforms use Persona-style verification under regimes like the UK's Online Safety Act; we are not asserting anyone dropped Persona over this leak.

The Thiel Thread

The money is worth naming plainly. Persona is backed by Founders Fund, Peter Thiel's venture firm, which led its major funding rounds. Thiel also co-founded Palantir, the surveillance-software company that works with US government agencies.[1][3]

That doesn't, on its own, prove the conspiracy some people want it to prove. What it does establish is the pattern: the same pool of capital sits behind both a marquee government surveillance contractor and the identity layer now wired into OpenAI and into federal financial-crime reporting. Draw your own line; just draw it with the facts above and not the ones we wish we had.

What You Can Do

Options are limited once your data is already in a system like this, but not zero:

  • Check platform privacy settings: a few services let you delete verification data after the check. Most don't.
  • Avoid verification when it's optional: not every age or identity check is legally required; some platforms do it voluntarily.
  • Prefer services that don't demand biometric ID: alternatives still exist for many use cases.
  • Use your data rights: under GDPR (UK/EU) and several US state laws, you can file a data subject access request to see what a vendor holds on you.
  • For UK users: the Online Safety Act is the engine pushing a lot of this. Reforming it is a political problem, not a settings toggle.

The real fix is regulatory: transparency requirements, data minimization, and hard limits on government data sharing for identity vendors. A 53MB government codebase left readable on a public URL is a decent argument that the current honor system isn't working.

Sources

  1. vmfunc (Celeste): The Persona / "Watchers" research writeup (primary source)
  2. Dataconomy: Is ChatGPT spying for the feds? The 53MB leak behind OpenAI's ID checks
  3. Cybernews: Persona leak exposes global surveillance capabilities
  4. ID Tech Wire: Exposed Persona frontend reveals extensive biometric surveillance stack behind age verification

Published: February 20, 2026