
Forty billion. That is how many bid requests a day BADBOX 3.0 pushed into the programmatic market at its peak, according to HUMAN Security, which announced the operation's disruption on Nov. 18. It is more than three times VASTFLUX, the previous record holder. HUMAN says it is a bigger peak than the last five years of its published investigations combined.
That is the press release number. The better number is in the full technical report.
At its height, BADBOX 3.0 accounted for more than 4% of all the connected TV traffic HUMAN's platform observed.
Sit with where that traffic was observed. HUMAN does not run an exchange. It scores bid requests that its customers send it: the SSPs, DSPs and platforms that pay for pre-bid fraud filtering. For HUMAN to see one of every 25 CTV requests coming out of infected TV boxes, those requests were already inside the pipes of the companies that sell and buy your CTV inventory.
Here is the line that belongs in every agency's next quarterly review. As of publication, not one SSP, not one DSP and not one verification company has said it found BADBOX 3.0 on its own.
The only company claiming the catch is the one that sells the catching.
What BADBOX 3.0 actually did
The hardware is the same species as the last two rounds. Cheap, off-brand, uncertified TV boxes running the Android Open Source Project, built in mainland China and shipped worldwide. HUMAN says they are not Android TV OS devices and not Play Protect certified. In the words of HUMAN's chief information security officer, Gavin Reid, if you bought one, it was compromised "before you plugged it in."
Once a box is on your Wi-Fi, a backdoor asks a command server for a list of modules and loads whatever the operators pick. The catalog lives on public code repositories, with backup copies on a second code platform, an operator file server and a free open-source CDN. The operators even run their own GitHub proxy through Cloudflare Workers, so blocking the repositories does not stop delivery.
The ad modules are the money. They load real video ad SDKs and play the ads in a window nobody can see. One system app on infected devices, com.android.systemsafe, loads three separate VAST video SDKs and runs CTV video campaigns in a transparent player parked outside the visible screen. Another poses as an app called "GoogleAd" and carries a method HUMAN's researchers found named, with real candor, showHideWindowAd.
Then comes the costume. The malware calls the same ad servers the official apps use, the same way the official apps do, and gets similar ad configurations back. The bid request names the legitimate app. Everyone downstream sees an ad opportunity inside a well-known CTV app. The real app never runs on the box. Most of the boxes probably do not even have it installed.
One detail in HUMAN's report deserves its own paragraph. The malware sends hardcoded values that let the ad server attribute the traffic to the threat actors.
Somebody's ad server knew which account to credit.
And while the TV plays ads to nobody, it rents your home to strangers. HUMAN found the same boxes enrolled in eleven residential proxy networks. Ten are commercial services: Pawns, Earn, Packet, GaGa Node, HoneyGain, Repocket, Bright, Cashraven, MystNodes and Infatica. The eleventh is PXYEDGE, which HUMAN says is IpMoYu under a new name, the proxy network run by the BADBOX crew and exposed in 2025. Nokia's researchers reached similar conclusions on their own. The report does not say whether the commercial services knew where those devices came from.
HUMAN ties the module catalog to TXCZ, a Chinese electronics manufacturer, through a stack of overlaps: a command server URL pattern that matches TXCZ's own device traffic, a TXCZ domain that serves payload downloads and the configurations for the hidden ads, a TXCZ launcher preinstalled at the factory with system permissions, modules that check for TXCZ-specific build properties before running, a repository account with TXCZ in its name, and a deleted file researchers recovered that turned out to be a distribution list of active TXCZ devices by model and firmware.
The tells were sitting in the bid stream
This is the part that should make every exchange's quality team uncomfortable. HUMAN lists the patterns that gave BADBOX 3.0 away. None of them required a malware lab.
A handful of devices. The heaviest traffic came from a small set of cheap, uncertified box models that sent requests almost exclusively to one family of apps. Real TVs spread their requests across many apps.
Residential proxies. For one impersonated app, about 40% of all requests and 80% of U.S. requests came from addresses flagged as residential proxy exits.
Shared addresses. Requests claiming to come from different apps tied to the operation came from the same IP addresses.
A machine rhythm. Volume surged at nearly the same hour every day. Nearly all traffic under one app's name reported a browser identifier from an internal test build, which no consumer TV would report.
The paperwork. The apps actually sending the requests do not appear in the impersonated apps' sellers.json or app-ads.txt files, where those files exist.
Every one of those signals sits in a bid request or a log file an SSP already holds. Device model. IP address. App bundle. Timestamp. User agent. Seller ID.
An exchange clearing billions of CTV requests a day can count how many arrive from one off-brand box model aimed at one app. It can see that 80% of a premium app's U.S. traffic is walking in through proxy exits. It can see a test-build browser string on a television. That is not forensic reverse engineering. That is a database query.
HUMAN says its own first signal, in mid-2026, was exactly this kind of thing: a steep rise in high-volume CTV requests claiming to come from a legitimate app, with an unusually high share of residential IPs. Its researchers tested the real app and the hardware it is built for and found nothing wrong with either. The anomaly showed up as traffic before anybody opened a box.
Who caught it, by their own account
The credit roll is short.
HUMAN found it, traced it and announced it. Google is quoted in the release, but from the Android security side, not the ad side. Shailesh Saini, who directs Android security and privacy engineering, thanked HUMAN and told people to buy Play Protect certified devices. Not a word about DV360 or AdX. Nokia is credited for its work on the proxy network, not the ad traffic.
That is the list.
We went through the public record for DoubleVerify, Integral Ad Science, Pixalate, The Trade Desk, Magnite, Index Exchange, PubMatic, OpenX, FreeWheel, Amazon and Google's ad platforms. We ran the same question through Perplexity and ChatGPT as a cross-check. Nothing. No statement of independent detection. No blog post. No "we flagged this in the spring."
The absence of a statement is not proof of a miss. Some platforms may have filtered this quietly. Several license HUMAN and filtered it through HUMAN.
That is the point. If your fraud detection is HUMAN, then HUMAN caught it, and you are paying a toll to stand in the same line as everybody else.
And these are the companies with the loudest quality pitches in the business. Index Exchange wrote on HUMAN's own blog in 2020 that its HUMAN partnership let it pursue being "the cleanest exchange for buyers and sellers." Jeff Green is quoted in HUMAN's sales material saying The Trade Desk partnered with HUMAN "to scan every single impression offered through our platform before we ever make a bid on it."
Forty billion requests a day went into that scanner. The scanner's owner is the only one who has said so.
This is the third time
Nobody gets to call this a surprise. The family has a calendar.
October 2023: HUMAN discloses the original BADBOX: backdoored off-brand boxes, more than 74,000 infected devices, and an ad fraud module called PEACHPIT that HUMAN put at more than 4 billion bid requests a day at its peak.
Late November 2023: Within weeks of that disclosure, the older BADBOX 3.0 repository is created.
March 2025: BADBOX 2.0. More than a million infected devices at first count. At peak, HUMAN says, 750 million fraudulent bid requests a day.
June 2025: Reid tells The Register, "we expect there will be a Badbox 3." He adds that the malicious traffic "often escapes normalized detections."
July 2025: Google sues the BADBOX 2.0 operators, alleging more than ten million infected devices.
January 2026: The old repository gets its last update the same day a new one is created. HUMAN calls it a planned migration.
March 15, 2026: pxyedge.com is registered.
Mid-2026: HUMAN sees the surge in impersonated CTV traffic.
July 2026: HUMAN discloses NewsJunkie, a separate CTV spoofing ring running up to nearly 2 billion invalid bid requests a day per seller.
Nov. 2, 2026: Google dismantles XCPE, a backdoor built into more than a million TV boxes, with HUMAN among the partners. HUMAN says that takedown hit several command servers running parts of BADBOX 3.0.
Nov. 18, 2026: BADBOX 3.0, at 40 billion a day.
From 750 million a day to 40 billion a day. That is roughly 53 times bigger.
The infrastructure behind it opened its doors within weeks of the first public report and stayed open for almost three years. The industry had the playbook since 2023. It had the hardware profile, the proxy network's name, the crew's habits and a public warning from the man who found it.
The supply paths took the traffic anyway.
Email signup to keep reading. Part two covers why the pipes said yes, the outsourcing problem nobody puts in the media plan, what the verification companies were saying while this ran, the numbers HUMAN did not publish, and the seven questions to send your SSP and DSP this week.
Part two: Why the pipes said yes
Nothing in the programmatic paperwork checks whether a television is real. That is the whole story in one sentence, and the rest of this is detail.
The paperwork checks sellers, not screens
The industry's answer to spoofing was ads.txt, app-ads.txt and sellers.json. They answer one question: is this seller allowed to sell this app's inventory?
They do not answer the question that mattered here: is this app actually running on this device right now?
A bid request is a set of claims. The app bundle is a claim. The device type is a claim. The IP is real, but in BADBOX 3.0 the IP was a real house, because the box was in a real house. HUMAN's report says the senders did not appear in the impersonated apps' sellers.json or app-ads.txt files where those files existed. Where those files exist is doing a lot of work in that sentence. In CTV, plenty of them do not.
HUMAN's NewsJunkie report in July showed how the rest of the gap gets closed. That operation used newly created intermediate sellers that stretched the supply path, then funneled into a small set of persistent sellers. The later hops were often authorized in sellers.json. The earlier hops were not. The apps were almost always unauthorized in app-ads.txt. One local news app in that scheme, with 185 reviews, generated more than 42.2 billion bid requests in two months.
An exchange that cannot notice a 185-review local news app outrunning the networks is not doing quality control. It is doing accounts receivable.
CTV hands the fraud a blindfold to wear
HUMAN has said it plainly. CTV is among the hardest environments to protect because a substantial share of the inventory runs with no JavaScript, which means far fewer signals for a detector to read. On desktop and mobile, a verification tag can poke the environment. On a TV fed by server-side ad insertion, there is often nothing on the device to poke.
BADBOX 3.0 used both sides of that. On the infected boxes, the ads really did play, so a video really did start and a tracker really did fire. And in some of the same supply paths, HUMAN found spoofed SSAI traffic: requests claiming a stitching server inserted an ad into a stream on a viewer's TV, when the TV did not exist and no ad was ever loaded.
Then there is the address. Residential proxies are built to defeat the oldest filter in the book. Data center lists and IP reputation scores catch bot farms. They do not catch a box behind a family's cable modem, because that is not a bot farm. It is a family's cable modem. Augustine Fou has been arguing for years that residential proxies break IP-based detection. BADBOX 3.0 is the argument with a 40-billion-request exhibit attached.
Notice, though, that the tells HUMAN listed did not depend on the IP alone. Concentration by device model. One test-build browser string. Different apps sharing addresses. A daily spike at the same hour. These are pattern checks, and pattern checks are cheap when you already hold the logs.
The outsourcing problem
Here is how fraud defense actually works in programmatic, as opposed to how it works on a sales slide.
The Media Rating Council spelled it out to ADOTAT in October. Pre-bid fraud vendors hand over lists or probability scores, by file or by API. They cannot block a bid. The publisher, the SSP and the DSP make the serving decisions. The MRC also told us invalid traffic slips through when buyers bill on unfiltered impressions, when nobody measures it independently, or when evasive techniques go undetected even after the bid. It flagged long-tail CTV as one of the places filtration rates run high.
So the decision belongs to the platforms. The intelligence, in practice, belongs to a vendor.
HUMAN's own case study with Magnite, the largest independent SSP, says the quiet part. It describes the BADBOX 2.0 investigation as giving Magnite insights it could act on "thanks to HUMAN's early discovery and briefing." Read it again. The biggest independent sell side in the market describes its BADBOX response as a briefing it received.
That is not a scandal. It is a business model. HUMAN's publicly named platform relationships include The Trade Desk, Magnite, Index Exchange, Xandr, FreeWheel, GumGum, OpenX, LinkedIn, AdRoll, Seedtag, Primis, EXTE and Madhive. The industry has consolidated its fraud research into one company's threat team, and the platforms pay for the feed.
The trouble is what that does to everyone's incentives. If the vendor has not found it yet, the traffic clears. If the vendor has found it, the platform blocks it and says nothing, because saying something invites the question of how much went through first. And an SSP paid on volume does not get a bonus for being the first to throw away a few billion requests a day that look like premium CTV.
We have seen what happens when the vendor's coverage gets tested from outside. Adalytics documented ad tech platforms that publicly cited HUMAN pre-bid filtering serving ads to declared bots sitting in data centers. Those were bots that announced themselves. BADBOX 3.0's boxes announced nothing.
None of this is a knock on the Satori team. They did the work. The question is what everybody else charges for.
What the verification companies were saying while it ran
The verification companies were not silent this year. They were describing the weather without naming the storm.
DoubleVerify reported that CTV fraud schemes and variants rose 140% in the first quarter of 2026 from a year earlier.
Pixalate put global programmatic CTV spend at $6.2 billion in the first quarter of 2026 and flagged 24.9% of that traffic as invalid. It said U.S. CTV invalid traffic rose 32% in the first half, and blamed free-form bundle IDs, the sprawl of CTV devices and spoofed SSAI. Its July report on the highest-traffic fraudulent CTV bundle IDs found 80% were brand spoofing.
Brand spoofing. Spoofed SSAI. Device sprawl. That is BADBOX 3.0's job description, published while BADBOX 3.0 was running.
Neither company has said it identified this operation. To be fair, neither has said it missed it. As of publication, they have said nothing about it at all.
The numbers HUMAN did not publish
HUMAN's report is the best public account of this operation, and it still leaves a buyer without the numbers a buyer needs.
How many devices. No count. The only hard figure is CDN statistics for one repository's third backup route: 367,023 requests and about 482.7 GB in one month. HUMAN says the real infection count is "significantly higher."
How much money. No dollar figure. A bid request is not a paid impression. Forty billion requests a day tells you about the flood, not the bill. Nobody has published a win rate.
Which apps. The report says one stream carried the identity of a legitimate CTV channel app and others impersonated well-known apps. It does not name them.
Which sellers and exchanges. Not named, in the release or the report.
How long it paid. The infrastructure dates to late 2023. HUMAN's traffic signal dates to mid-2026. The gap between those two dates is where the money is, and nobody has filled it in.
Whose ad server. The malware sent hardcoded values so an ad server could credit the threat actors. Which ad server, and whose account got paid, is the most important open question in the report.
Seven questions to send your SSP and DSP this week
Put these in writing. Ask for answers in writing. A shrug on a call is not an answer.
Did BADBOX 3.0 traffic reach your platform, and how many requests a day at peak?
When did you start filtering it, and was that your own detection or HUMAN's briefing?
Which seller IDs carried it, and are any of them still active?
Did any of it clear and bill, and have buyers been credited?
Do you flag cases where one device model sends nearly all its requests to one family of apps?
What share of your CTV supply arrives from residential proxy addresses, and what do you do with it?
Before you sell a request as a named app, do you confirm the sender appears in that app's app-ads.txt and the sellers.json chain?
If the answer to number two is "HUMAN," ask the follow-up. What are we paying you for that HUMAN is not already doing?
Forty billion requests a day is not a needle. It is the haystack. The platforms selling "verified" CTV were standing in it, and the only one who said so was the company they pay to look.
How we reported this: This piece draws on HUMAN Security's Nov. 18, 2026 BADBOX 3.0 technical report and press release; HUMAN's earlier reports on VASTFLUX, BADBOX, BADBOX 2.0 and NewsJunkie; HUMAN's Magnite case study and published partner announcements; Google's statement in the release; Nokia Deepfield's public research; reporting by The Register and PPC Land; DoubleVerify and Pixalate research; Adalytics; and the Media Rating Council's written answers to ADOTAT in October 2026. On Nov. 18 we searched the public record and used Perplexity and ChatGPT search as cross-checks for any SSP, DSP or verification company claiming detection; none was found. HUMAN's figures have not been independently verified, a bid request is not a billed impression, and the absence of a public statement does not prove a platform missed this traffic, since some filter through HUMAN. Named companies have not yet been asked for comment; responses will run as dated updates. Pesach Lattin ran an ad network in the late 2000s and has known executives at several companies named here for years. ADOTAT is reader-supported. No paid placement in editorial.

