This website uses cookies

Read our Privacy policy and Terms of use for more information.

Troutman Amin

The Field Kit, Thursday November 19, 2026

Jeff Green has a line in HUMAN Security's sales material. The Trade Desk, he says, partnered with HUMAN "to scan every single impression offered through our platform before we ever make a bid on it."

Good. Then somebody has the scan. On Wednesday, HUMAN said an operation called BADBOX 3.0 pushed roughly 40 billion bid requests a day into the programmatic market at its peak, out of cheap TV boxes playing video ads in a window nobody could see. At its height it was more than 4% of all the connected TV traffic HUMAN's platform observed.

Every single impression got scanned. As of publication, not one SSP, DSP or verification company has said what its scan found.

Yesterday we did the autopsy: how the boxes worked, why the pipes said yes, and who has claimed the catch. That list has one name on it, and it belongs to the company that sells the catching. Today is the part you can use. Eighteen questions, sorted by who you're asking. Five checks you can run on your own log data. A 14-day rule your vendors would rather you didn't read. Four clauses. One email.

One caution before you start typing. We don't know how much of this traffic The Trade Desk, Magnite, Index Exchange or anybody else saw, bid on or billed. HUMAN hasn't named a single seller or exchange. You don't know either. That is the whole problem, and it's the one piece of this story you can fix by Friday.

The Sentence You're About to Receive

"Our customers were protected." HUMAN said its version on Wednesday. Expect your SSP to say it by Friday and your DSP to forward it by Monday.

It's a great sentence. It has a subject, a verb and no numbers. Take it apart.

  • Protected since when? HUMAN says its first traffic signal arrived in mid-2026. The infrastructure behind BADBOX 3.0 dates to late November 2023. A filter that switches on in July protects you from August.

  • Protected by whom? The Media Rating Council told ADOTAT in October that pre-bid fraud vendors hand over lists and probability scores. They cannot block a bid. The platform decides what to do with the score. "Protected" means somebody acted on it. Ask who, and at what threshold.

  • Protected on which requests? HUMAN scores what its customers send it. Supply that reached your campaign down a path that never called HUMAN was scored by whoever that path did call. Possibly nobody.

  • Which customers? HUMAN's customer is the platform. You are the platform's customer. That puts you two contracts away from the word "protected."

Protection is a date, a threshold and a dollar amount. Anything else is a mood.

Four Companies Have Your Receipt. You Have an Invoice.

Every CTV impression you bought left a record in at least four places.

  • The SSP logged the request: the app, the device, the IP address, the seller and the path it took to get there.

  • The DSP logged the bid, the win and the price.

  • The verification vendor logged whatever its tag or its server could see.

  • The agency has a login to all three.

You have a PDF with a total on it. Everything in this kit is a way of turning that PDF back into the records it was made from.

One Check You Can Run Before Lunch

You don't need a malware lab. HUMAN's report lists the tells, and one of them is sitting in a public text file.

The apps that actually sent BADBOX 3.0's requests did not appear in the impersonated apps' app-ads.txt or sellers.json files, where those files existed. So go look.

  1. Open your last CTV delivery report and pull the top ten app bundles by spend, with the exchange and seller ID for each. If your report doesn't show seller IDs, that's your first finding.

  2. For each app, find the developer's website on its app store listing and open /app-ads.txt on that domain.

  3. Search the file for the exchange's domain and the seller ID that sold you the impression.

  4. Sort what you find into three piles: listed as DIRECT, listed as RESELLER, not listed.

"Not listed" means you bought a famous app from somebody the famous app has never heard of. "No file at all" is its own pile, and in CTV it's a big one. Pixalate's count for the first quarter of 2026: 53,000 apps listed on Roku and 13,000 with app-ads.txt files. On Amazon Fire TV, 116,000 apps and 18,000 files. Apple TV, 26,000 and 3,000. Most of the CTV app universe has never published a list of who's allowed to sell it.

Either way you now hold a list of seller IDs and a specific question, which is more than the press release gave you.

The 14-Day Rule Nobody Mentions

The Media Rating Council's invalid traffic standards have a section most buyers have never opened. It's Section 4.5 of the June 2020 standards, titled "Requirement for Backward-Looking Assessments and Correction."

Here is what it says, in plain English. When an MRC-accredited measurement company finds a new type of invalid traffic, it has to look backward. The review must cover at least the 14 days before the discovery date, and campaigns that run for months are supposed to be evaluated over the whole campaign. If the review turns up material errors, the company "has a responsibility to inform customer users of the data." Buyers and sellers both. The same standards treat anything at 5% or more of reported activity as material.

Two things follow.

  • There's a discovery date, and the clock runs backward from it. HUMAN says its first BADBOX 3.0 signal showed up in mid-2026. Every accredited company that changed its filtering because of this operation owes somebody a look-back.

  • You can sign this away, and you might have. The MRC lets a customer waive the look-back, but only formally in the campaign terms, and it says the waiver has to be an exception, not standard boilerplate.

So ask two questions: Did you run the Section 4.5 review? Does my contract say I waived it? If the answer to the second one is yes, find out who signed it.

Three Questions to Send Today

Put them in writing. Ask for the answers in writing. Yesterday we gave you seven questions. These are the three that matter most, with the answer you're going to get and the reply that makes that answer cost something.

1. How many BADBOX 3.0 requests reached your platform at peak, and on what date did you start filtering them? What they'll say: "We work closely with HUMAN, and our clients were protected." What you say: "That's a relationship, not a number. Give me requests per day at peak, the date filtering began, and the count that cleared before that date."

2. Was that your detection, or HUMAN's briefing? Magnite's own HUMAN case study credits its BADBOX 2.0 response to "HUMAN's early discovery and briefing." That's honest. It also tells you who did the discovering. What they'll say: "We take a multi-layered approach that combines proprietary systems with best-in-class partners." What you say: "Which layer flagged this one, and on what date? If the answer is the partner, tell me what your layer costs me."

3. Did any of it bill to my account, and where is the credit? A bid request is not an impression, and most of 40 billion a day never became an invoice. Some of it had to, or there was no business. What they'll say: "We have no evidence of material impact on your campaigns." What you say: "Then you ran the query. Send it: impressions on the affected seller IDs and app bundles, by day, June 1 through Nov. 18, with what I paid."

The Mic Drop

The Trade Desk says every single impression gets scanned before it bids. Index Exchange wrote in 2020 that its HUMAN partnership let it pursue being "the cleanest exchange for buyers and sellers." Magnite thanked HUMAN for the briefing.

Fine. That's three customers of one scanner. The scanner's owner says it watched 40 billion fake requests a day walk into connected TV, and says its customers were protected.

So here is the industry's position in full: the fraud was the biggest ever disclosed, everyone was protected, and nobody can tell you how much they stopped. Pick any two.

You aren't paying a verification fee, a platform fee and an exchange fee for a press release. You're paying for the log file. Go get it.

Below the line for members: two clocks that are already running, five checks to run on your own log data, fifteen more questions for your SSP, your DSP, your verification vendor and your agency, how to size a credit claim and what form it comes in, a paste-ready email, four clauses for the next renewal, and how to read the answers.

This kit builds on ADOTAT's Nov. 18 report and draws on HUMAN Security's BADBOX 3.0 technical report and press release, HUMAN's July NewsJunkie report and 2023 supply chain research, its Magnite case study and published partner statements, Section 4.5 of the Media Rating Council's June 2020 Invalid Traffic Detection and Filtration Standards, the MRC's written answers to ADOTAT in October 2026, Google's DV360 help and developer documentation, Pixalate's Q1 2026 CTV supply chain report, April 2026 CTV fraud types report and Q1 2024 supply chain verification report, DoubleVerify research, and The Register. Research was cross-checked with ChatGPT and Perplexity, and every figure was checked against its original document. HUMAN's figures have not been independently verified, and a bid request is not a billed impression. Nothing here is a finding that The Trade Desk, Magnite, Index Exchange, Google or any other named company carried, bought or billed BADBOX 3.0 traffic; HUMAN has not named sellers or exchanges.

logo

You've read this far. Read the rest with us.

Members fund the reporting and read it first. No advertisers, no sponsored anything, just the people who want this work to exist.

Join ADOTAT+

Already a subscriber? Sign in

What members get:

  • Every investigation, including the ones with a lawyer's fingerprints on them. Those are the good ones.
  • The Field Kit. Because "do your own diligence" is useless advice without the actual checklist.
  • Their side, unedited. You get the full statement before I take it apart. Judge for yourself which of us is being unfair.
  • No paid editorial. Not one word, not ever. If it's in the copy, nobody paid to put it there.