HUMAN says BADBOX 3.0 peaked at 40 billion bid requests a day and links it to TXCZ; Google named Videostrong for XCPE
HUMAN's Satori team says BADBOX 3.0 peaked at roughly 40 billion bid requests a day and links its module catalogue to the manufacturer TXCZ. The report counts no infected devices and no dollars, and Google named a different maker, Videostrong, for the related XCPE botnet on 2 November.
By Parminder Kumar Sharma · LinkedIn · · 23 min read
40 billion bid requests a day is about 463,000 a second, and HUMAN counts no devices (derived)
HUMAN's Satori team published its BADBOX 3.0 report on 18 November 2026 (its own stamps: published 13:35 UTC, modified 19:06 UTC; this site saved the page on 21 November at about 15:56 UTC). It calls BADBOX 3.0 "the largest ad fraud operation by bid volume ever publicly disclosed" and says that at its height the operation accounted for "roughly 40 billion bid requests a day" and "more than 4% of all CTV traffic observed by the Human Defense Platform". Spread evenly over a day, 40 billion is about 463,000 bid requests a second (derived, rounded from 462,963: 40,000,000,000 divided by 86,400 seconds; HUMAN says volume surged at nearly the same hour each day, so this is an arithmetic illustration and not a measured rate).
What that does not establish. The unit is a bid request, an offer of one ad slot sent to buyers. It is not a device and not a dollar. The report prints no count of infected BADBOX 3.0 devices, no loss figure, no date for the peak, and no names for the apps whose identities the traffic borrowed. Its one delivery number is a floor for a single repository, a single backup route and a single month: 367,023 requests. The report links the module catalogue to TXCZ, "a Chinese electronics manufacturer", and prints no reply from TXCZ. Google named a different maker, Videostrong, for the related XCPE botnet on 2 November, and its post does not mention BADBOX 3.0.
HUMAN sells Ad Fraud Defense and related products, and the report says customers using them are protected from the effects of BADBOX 3.0 ad fraud. That is the publisher's selling interest. Nobody pays this site, and it shares no vendor's interest. The report lists eight authors and BleepingComputer names its own; this site names none. A BleepingComputer article of 21 November (10:17), a secondary, was used only to find the primary. Its headline says ad requests, where the primary says bid requests.
What the headline number is, and what it is not
The report opens with three figures: 40 billion bid requests a day at peak, more than 4% of the CTV traffic its platform observes, and more than three times the volume of the previous record holder, VASTFLUX. It adds that this is "a larger peak than the last five years' worth of published Satori investigations combined". The table sets each beside what it leaves open.
The report's headline figures and what each leaves open (HUMAN report, saved 21 November 2026)
| Figure as printed | What it is | What it does not give |
|---|---|---|
| About 40 billion bid requests a day, at its height | A count of the bid requests HUMAN observed and associated with the operation | Infected devices, money lost, the day of the peak, or how much of the 40 billion is the spoofed server side stream |
| More than 4% of CTV traffic observed | A share of what the Human Defense Platform sees, at peak | The size of that traffic, or the market's. HUMAN does not say the 40 billion and the 4% are the same moment |
| More than three times VASTFLUX | A ratio to the previous record holder | VASTFLUX's own number in the text. HUMAN's Figure 1 compares operations by daily bid requests, but it is an image whose numbers this site did not read |
| A larger peak than five years of published Satori investigations combined | A comparison with HUMAN's own earlier reports | Which investigations, or their figures |
Two bounds follow from the printed ratios, both derived by this site from rounded words. If "more than three times" is read literally, VASTFLUX's peak was under about 13.3 billion a day (derived: 40 billion divided by 3). If the 40 billion and the 4% describe the same moment, the CTV traffic HUMAN observed then was under about 1 trillion requests a day (derived: 40 billion divided by 0.04 is 1,000 billion). HUMAN does not say they are the same moment. These bound; they do not measure.
A second 40 appears in the report and is easy to confuse with the first: for one abused app identity, about 40% of requests overall and 80% of US requests came from addresses flagged as residential proxy exits. That is a share of one identity's requests, not part of the 40 billion count.
How the box, the module list and the two payloads fit together
HUMAN describes the box as low cost, off brand, uncertified hardware running the Android Open Source Project. It says these are not Android TV OS devices and not Play Protect certified. The investigation began with a backdoor packaged as a system application with two hardcoded endpoints: one returns the list of modules to load, the other handles logging. The command server's answer names each module and tells the device where to fetch it.
The modules come from what HUMAN calls a catalogue. Each has a primary location on public code repositories and backups: a second code platform, a file server the operators control, and for some payloads a mirror on a free open source content delivery network that keeps files up even if the source repository is removed. The operators also run their own copy of an open source GitHub proxy so that network level blocking of the repositories does not stop the droppers. The version counter on the module manifest had reached the mid 500s by early 2026, which HUMAN reads as sustained development, and it says antivirus coverage of the newer modules was effectively absent at the time of analysis.
Most modules are residential proxy modules and the droppers that install them. The ad fraud modules are fewer, but HUMAN says they drive the largest part of the volume. They load real video ad software development kits, including kits belonging to the abused apps, and play the ads in a window the owner cannot see: in one example a transparent player positioned outside the visible screen. In HUMAN's lab, one sample's hidden ad code path was complete, but the server "did not return a configuration", which HUMAN puts down to device fingerprinting or server side targeting. That sample was not seen serving ads in the test.
The abused apps never run on the infected devices, and most of those devices likely do not have them installed. The malware contacts the same ad servers the real apps use, in the same way, so the resulting bid request names the legitimate app and everyone downstream sees an ad opportunity inside it. HUMAN also says it tested the legitimate app it first noticed, and the hardware it is designed for, and "found no malicious behavior". That is HUMAN's finding that the impersonated app was not the source of the traffic.
The three channels, read word by word
A reader might expect the report to split the 40 billion into three named channels. The saved page says the volume "came through three channels" but labels two, then makes a statement about the whole. The table keeps the page's own words in order.
HUMAN's description of how the 40 billion arrived, as printed (report saved 21 November 2026)
| Position in the text | What the page says | What it does not say |
|---|---|---|
| First | A set of malware bearing Android apps disguised as system components, whose bid requests HUMAN assesses as "entirely invalid" | How many of the 40 billion, or whether this was the largest part |
| Second | A stream carrying the identity of a legitimate CTV channel app | How many of the 40 billion, or which app |
| Third (not labelled) | "The vast majority of that 40 billion, however, carried spoofed, misrepresented, or fabricated traffic attributes." | Whether this describes a third channel or the whole count, and how it splits from the first two |
| Separate section | Spoofed server side ad insertion needs no infected device: a request claims that a server inserted an ad into a stream on a viewer's device, but the device does not exist and no ad was loaded | Whether this stream is inside the 40 billion, and how large it is |
So the one sentence that carries the weight, that the vast majority was spoofed, misrepresented or fabricated, is a statement about the 40 billion. It does not say what share needed a BADBOX 3.0 box. A reader can take two things from it without over reading: the headline is not a count of ads that played on infected televisions, and the page gives no split between box generated and server generated requests.
The only delivery number is a floor for one repository, one route and one month
HUMAN uses the usage statistics of the open source CDN mirror as "a possible low end of delivery figures". Over one month the mirror for a single repository served 367,023 requests and 482,659,426,190 bytes. Dividing, 482,659,426,190 / 367,023 = 1,315,066 bytes, about 1.315 MB a request (derived; HUMAN prints about 482.7 GB and about 1.31 MB, which it says matches the size of the modules). HUMAN is plain about the limits: "These are requests rather than unique devices, and some come from researchers and scanners." It adds that "these numbers are only for one repository, and only of the third backup delivery mechanism", and that "The actual number of BADBOX 3.0 infections can be presumed to be significantly higher".
That is the report's whole answer on how many boxes. It is a count of module downloads, on one route of four, for one of at least two repositories, over one month. It cannot be compared with 40 billion bid requests a day: they are different units over different periods. The report does not say how many boxes are infected, and this site does not estimate it.
Who is behind it: five links HUMAN calls links
HUMAN says Satori researchers "linked the BADBOX 3.0 catalog to TXCZ" through "several independent connections". It numbers four and adds a fifth with "Finally". The table gives each as the page words it, and what that item does not show.
HUMAN's connections between the BADBOX 3.0 catalogue and TXCZ (report saved 21 November 2026)
| Connection | What HUMAN prints | What it does not show |
|---|---|---|
| 1. Command server | The server that tells devices which modules to fetch uses a domain HUMAN prints in its indicator table, and its traffic follows a URL structure. HUMAN says a TXCZ owned and branded website shows the same nomenclature in its own device communications, and that an earlier module connected directly to the address behind it. | A shared naming pattern between two services. The page does not say who built either, or that one operator runs both |
| 2. A TXCZ domain | A TXCZ domain offers primary and secondary routes for devices to download payloads. A subdomain is where one ad fraud sample requests its ad configurations. | In HUMAN's own test the server returned no configuration for that sample, so the report shows the request, not the ads it would have returned |
| 3. TXCZ's own software | A launcher with system level permissions comes preinstalled on TXCZ manufactured devices, which HUMAN reads as factory installed. A module checks for TXCZ specific build properties before it proceeds. | That the launcher itself is malicious: the page presents it as the device a module is written for, and the module's check as showing intent to run on those devices |
| 4. The repositories | One repository sits under an account name incorporating TXCZ. Deleted files recovered from version history included a distribution list of active TXCZ devices by model and firmware version. | What the list was for, who wrote it, or how many devices it holds |
| 5. An actor | HUMAN has identified "a threat actor maintaining the list of modules" and linked that actor to TXCZ. | The nature of the link. The page gives no detail, and this site names no one |
"Linked" is the report's word, and it is a friendly one. It is not "operated", "built", "sold" or "knew", and the report does not use those words about TXCZ. It prints no confidence level for the TXCZ link (it uses the word confidence once, for its assessment that PXYEDGE is IpMoYu), no reply from TXCZ, and no count of TXCZ devices. This site did not look for a reply, because it fetched nothing. The five items fit more than one account: a maker that built the catalogue, a maker whose devices and servers someone else used, or a supplier arrangement the report does not describe. This site does not say which. The report does not choose among them.
Google's separate post names a different maker. On 2 November 2026 Google's Threat Intelligence Group said it assessed "with high confidence" that the XCPE backdoor and botnet were "intentionally developed, deployed, and operated for financial gain" by Videostrong Technology Co., an Android original equipment and design manufacturer, that Videostrong monetised the backdoor by selling access to criminal affiliates, and that more than 1.3 million set top boxes were enrolled. This site's earlier briefing covers that post. HUMAN says it took part in the Google led disruption, which it says found more than a million infected devices in more than 200 countries and territories. HUMAN calls XCPE "distinct from, but related to" BADBOX 3.0 and says "the XCPE disruption included multiple C2 domains identified as operating portions of BADBOX 3.0".
What the saved Google post says in return: it lists HUMAN Security among the partners it worked closely with, and it does not mention BADBOX 3.0, TXCZ or Satori. It prints the word BadBox 3 times, as earlier Google enforcement and as a past family that two of the modules it describes were tied to. HUMAN's report does not name Videostrong. Both reports name MoYu, Google as a backdoor module it saw on XCPE boxes and HUMAN as lineage in its catalogue, and both name HoneyGain and GaGa Node among the proxy services. Those overlaps are in the text. Whether the two makers are connected, and whether any device is counted in both reports, is stated by neither. Sixteen days separate the two reports (derived: 2 November to 18 November).
What the traffic looks like to a buyer, and four tells
HUMAN says that to buyers the traffic looks like "premium CTV inventory": a well known app, a residential address and a full video ad. The sending apps have no business relationship with the apps they claim to be, and they do not appear in the claimed apps' sellers.json or app-ads.txt files where such files exist. Those two files are where an app publisher lists who is allowed to sell its inventory. A name in a bid request is a label, and a label is not a control until something checks it against those lists.
The four tells HUMAN prints, and what each asks of a buyer (tells from the report; the right column is this site's suggestion)
| Tell (HUMAN) | What the report says | A check a buyer could run (this site's suggestion) |
|---|---|---|
| A handful of devices | The heaviest traffic came from a small set of inexpensive, uncertified AOSP device models that sent requests almost exclusively to one family of apps. Ordinary CTV devices spread requests across many apps | Group requests by device model and look for models that name only one family of apps |
| Residential proxy addresses | For one abused app identity, about 40% of requests overall and 80% of US requests came from addresses flagged as residential proxy exits | Ask your supply path for the share of impressions from residential proxy addresses, by app |
| Shared addresses | Bid requests claiming to come from different apps tied to the operation shared the same IP addresses | Look for one address presenting as several unrelated apps |
| A machine rhythm | Volume surged at nearly the same hour every day, and nearly all traffic under one abused app's identity reported a browser identifier from an internal test build, which consumer TVs would not report | Plot requests by hour and look for a daily surge, and for a user agent consumer TVs would not send |
None of the four is a device test. All four are read off the traffic itself, which is why HUMAN says its detection targets the traffic from compromised devices rather than everything that carries the abused apps' names. Because the report does not name the abused apps, a buyer cannot check a specific app against it.
Eleven networks, one rebranded, and a lineage with counts that stop
The second payload enrols the box in one of eleven residential proxy networks. HUMAN names ten as commercially available services: Pawns, Earn, Packet, GaGa Node, HoneyGain, Repocket, Bright, Cashraven, MystNodes and Infatica. It marks four of them, Earn, Packet, GaGa Node and HoneyGain, as networks targeted for disruption in the XCPE takedown. The eleventh, PXYEDGE, it says is run by the threat actors themselves and assesses to be IpMoYu, the proxy network from BADBOX 2.0, under a new name. HUMAN says the owner pays for the bandwidth and someone else sells access to the address. A named service is HUMAN's finding about where modules enrolled boxes, not a finding about what that service knew.
PXYEDGE figures as HUMAN prints them, with this site's derived values (report saved 21 November 2026)
| Item | As printed | Derived by this site |
|---|---|---|
| Domain registered | 15 March 2026, with no advertising or promotion | 248 days before the report's 18 November publication (15 March to 15 November is 245 days, plus 3) |
| Nodes advertised | About 200,000 available nodes | None |
| Unique connections at peak | About 340,000 across all monitored services, from a publicly accessible monitoring server | 1.7 times the advertised nodes (340,000 / 200,000) |
| Server clusters | A numbered convention reaching at least 43 clusters | None |
| Independent view | HUMAN says Nokia independently reached similar conclusions | This site did not read the Nokia analysis |
The BADBOX family as HUMAN states its counts (report saved 21 November 2026)
| Chapter | Devices, as HUMAN states | Lead payload |
|---|---|---|
| BADBOX, 2023 | More than 74,000 infected devices | Advertising fraud named PEACHPIT, with a firmware backdoor based on Triada |
| BADBOX 2.0, 2025 | More than 10 million devices across 222 countries and territories | Click fraud, two kinds of advertising fraud and residential proxy services, through a backdoor named BB2DOOR |
| BADBOX 3.0, 2026 | Not stated. The headline figure is 40 billion bid requests a day | Hidden video ads and eleven proxy networks, through a module catalogue |
The pattern in the third table is the point to carry away. The first two chapters were reported in devices; this one leads with a request count. HUMAN does say the infection count "can be presumed to be significantly higher" than its delivery floor, so the box count is missing from the report and not from the operation. HUMAN also says several modules belong to vo1d, a malware strain first disclosed by Dr.WEB in 2024, which it says reinforces the earlier association it drew with BB2DOOR.
What HUMAN says was done, and what it does not say
The report says HUMAN took part in the Google led XCPE disruption, and that its customers are protected: Ad Fraud Defense and Ad Click Defense against the ad fraud, and Sightline Cyberfraud Defense and AgenticTrust against account takeover, fake account creation and carding routed through residential proxies. On the repositories, the only statement is the caption of its Figure 4, which calls a BADBOX 3.0 related repository "now taken down". The body text says nothing more, and it does not say which repository, who removed it, or when. BleepingComputer's text says one malware repository was taken down; that is a secondary repeating the caption's idea, not extra evidence.
HUMAN expects the operators to respond: "Adaptation should be expected", it writes, because adaptation has followed every prior disclosure in this family. Its conclusion is that disrupting an operation and dismantling the conditions that produced it are different achievements. Not stated: what Google, the code hosting platforms or the content delivery network did about the repositories and mirrors, whether the other repositories and the catalogue's other routes are still up, and whether the XCPE takedown cut any BADBOX 3.0 device off from its operators.
What HUMAN states, and what it does not
Stated and not stated in HUMAN's report (saved 21 November 2026)
| Question | What the report states | What it does not state |
|---|---|---|
| How big is it? | Roughly 40 billion bid requests a day at its height; more than 4% of CTV traffic HUMAN observes | Devices, dollars, the date of the peak, the split among channels |
| Who is behind it? | The catalogue is linked to TXCZ through five connections | That TXCZ operated or knew; any reply from TXCZ; a confidence level |
| Is it the same as XCPE? | XCPE is distinct from but related to BADBOX 3.0; the XCPE disruption included command domains operating portions of it | Videostrong, or any link between that maker and TXCZ |
| Which apps were abused? | Legitimate CTV apps whose identities the requests borrowed; the apps are not running on the boxes | Their names |
| How many boxes? | Nothing counted. One repository's CDN mirror served 367,023 requests in a month | Unique devices, and any total |
| What happened to the infrastructure? | One repository is captioned as taken down | Which one, by whom, when, and the status of the rest |
| Is it over? | HUMAN expects adaptation | Whether BADBOX 3.0 traffic continues today |
Where the pages differ, with nothing resolved
Differences between the saved pages (read 21 November 2026)
| Question | One page says | Another page says |
|---|---|---|
| What is the unit? | HUMAN: bid requests a day | BleepingComputer's headline: ad requests; its own text says bid requests |
| Is a repository down? | HUMAN's Figure 4 caption: now taken down; the body text is silent | BleepingComputer's text states one malware repository was taken down, without a source beyond HUMAN |
| What did the XCPE disruption do to BADBOX 3.0? | HUMAN: it included multiple command domains identified as operating portions of BADBOX 3.0 | BleepingComputer: it included taking down multiple command domains used by BADBOX 3.0. Google's post does not mention BADBOX 3.0 |
| Which maker? | HUMAN: TXCZ, linked, for the BADBOX 3.0 catalogue | Google: Videostrong, assessed with high confidence, for XCPE |
| Which device count? | HUMAN: none for BADBOX 3.0; more than a million for XCPE | Google: more than 1.3 million boxes for XCPE, and at least 10 million more reached by affiliate command domains |
| Certified or not? | HUMAN: BADBOX 3.0 devices are not Play Protect certified | Google: Videostrong used obfuscation to exploit a gap in Google's tests and load code after the first boot, and Google banned it from building certified Android devices |
The names and domains HUMAN prints
HUMAN's report prints three indicator tables of domains and addresses, and names in its text that a defender can search for. This site prints the domain names exactly as HUMAN's tables print them (defanged with square brackets), grouped as the report groups them, and the names from the prose. It prints no IP address and no port. It counts them instead: the three tables hold 33 rows (7, 8 and 18), of which 20 are addresses (1, 2 and 17) and 13 are domains (6, 6 and 1), all derived by this site by counting. One domain row carries a port number in HUMAN's table, which this site leaves off. Defenders can take the address rows from HUMAN's own table. Never visit or resolve any of these: search your DNS and proxy logs for lookups, and leave the hosts alone.
Command and payload hosting. HUMAN's first table. The first four are one family of names: the module list and logging endpoint the backdoor contacts, a payload host, the GitHub proxy and the parent domain. The table also lists one legacy command server address, not printed here.
api.drgdj[.]top Module list and logging endpoints contacted by the backdoor; module-loading backend contacted by ad fraud samples
file.drgdj[.]top Operator-controlled payload host
gh.drgdj[.]top Self-hosted GitHub proxy used to bypass network blocking
drgdj[.]top Parent domain
api.pixelpioneerss[.]com Module C2
bur.thedynamicleap[.]com Module C2
Ad fraud infrastructure. HUMAN's second table. The last row is a domain of the manufacturer HUMAN links to the catalogue. A lookup for it shows that a device is talking to a TXCZ service, which is where the launcher HUMAN describes would connect (this site's reading). It does not show that a hidden ad module ran. The table's other two rows are addresses, not printed here.
adv.speed25[.]top Ad behavior management for com.speed.adv samples
ad3g[.]com Ad fraud and tasking C2
moyu88[.]xyz Ad fraud and tasking C2
ziyemy[.]shop Ad fraud and tasking C2
terrtest.oss-cn-hangzhou.aliyuncs[.]com Cloud storage endpoint contacted by ad fraud modules
api.shzhtxcz[.]com Serves hidden ad configurations (/videoadmin/api/advertisement/getByModel) and APK download URLs
PXYEDGE. HUMAN's third table lists the storefront domain and seventeen addresses with roles such as command server, relay, database, proxy and the numbered myic servers. Only the domain is printed here.
pxyedge[.]com Storefront domain, registered March 15, 2026
Names in the text. These are the build properties, package and class names, file name and URL components HUMAN prints. A managed device inventory or a package list can be searched for the package names. The build properties tell you a device is TXCZ built. HUMAN prints no hashes, no log lines and no YARA rules.
Build properties a module checks before it proceeds:
ro.txcz.model
persist.sys.txcz.adopen1
Launcher preinstalled on TXCZ manufactured devices:
com.shzhtxcz.txbox
Module class names imitate Android system packages:
com.google.system.*
Example system app that loads three VAST video ad SDKs:
com.android.systemsafe, with advertising package com.speed.adv
Example app that presents itself as GoogleAd:
activity com.google.AdActivity
Module file that references earlier MoYu classes:
admoyu_dex.jar (com.ast.sdk, com.m.s.a.admy, com.mo.gd.inter.SDKMain)
URL structure components HUMAN points to on the command server and a TXCZ site:
/dd/cc2 and /dd/cc
Ad configuration path HUMAN prints in its indicator table:
/videoadmin/api/advertisement/getByModel
What to do, in order
First HUMAN's advice, then this site's suggestions, marked as such.
Take this with you
Actions, in the order worth doing
- HUMAN's advice to end users: choose devices from familiar brands, check whether a device is Play Protect certified, download apps only from official marketplaces, and retire a device that behaves strangely.
- This site's suggestion: inventory the off brand streaming and TV boxes on office, hotel and home networks you are responsible for, and keep them on a separate network segment from the machines that matter.
- This site's suggestion: search DNS and proxy logs for lookups of the 13 domains in HUMAN's tables (and take the 20 address rows from HUMAN's page). A hit tells you which device to examine. Treat a hit on the TXCZ domain as a sign of a TXCZ built device, not proof of a hidden ad module.
- This site's suggestion, for buyers and publishers: compare the app named in a bid request with the claimed app's sellers.json and app-ads.txt, and look for HUMAN's four tells in your own logs.
- This site's suggestion: ask your supply path what share of its CTV impressions arrive from residential proxy addresses, by app, and what it does with traffic that fails the sellers.json check.
- This site's suggestion: treat Play Protect certification as one signal. HUMAN says the BADBOX 3.0 devices are uncertified, but Google's XCPE post says its certification tests had a gap that a maker exploited, so a certified label is no longer proof by itself.
- This site's suggestion: where a box cannot be updated, cannot be examined, or cannot be shown to be what it says, retire it. HUMAN expects the operators to adapt, so a block list of today's domains is a short term control.
The question the 40 billion leaves open
HUMAN's own description is that this traffic looked like premium CTV inventory to buyers: a known app, a household address, a full video ad. Its tells are in the traffic, and its fix, for its customers, is detection after the request arrives. If 40 billion requests a day looked like premium inventory, what in your buying path would have said otherwise: a sellers.json check, a residential address share, a daily surge, or nothing at all?
Key facts
Sources
- PrimaryBADBOX 3.0 Hides the Largest Ad Fraud Operation Ever Disclosed Inside Infected TV Boxes: published 18 November 2026 13:35 UTC, modified 19:06 UTC, saved 21 November at about 15:56 UTC; the primary, read in full (its figures are images and were not read)HUMAN Security (Satori Threat Intelligence report)accessed 2026-11-21
- PrimaryShutting the Door on the XCPE Botnet, 2 November 2026, as saved for this site's earlier briefing and read again: the maker Videostrong, the assessment with high confidence, 1.3 million set top boxes, partners including HUMAN SecurityGoogle Threat Intelligence Group (Google Cloud Blog)accessed 2026-11-21
- Reported byBADBOX 3.0 botnet returns with record-breaking 40 billion daily ad requests, 21 November 2026 10:17, saved 15:55 UTC: secondary report, used only to find the primary and to compare its headline unit and its statements on the repositoryBleepingComputeraccessed 2026-11-21


