On November 18, 2026, HUMAN Security announced that its Satori team disrupted BADBOX 3.0, describing roughly 40 billion bid requests a day at peak and hidden video modules on cheap Android TV boxes. The trade press framed volume and residential proxies. The technical report is a playback story: malware loads legitimate VAST video SDKs, requests the same ad configurations real apps use, and renders in a transparent window positioned outside the visible screen.
That split matters to anyone who signs off on CTV tags before money moves. Pre-bid filters score the OpenRTB request. Verification vendors score the session that fired. Neither step substitutes for asking whether the player surface that executed the VAST was visible on the glass the household watches. BADBOX 3.0 is evidence that the two checks can diverge while the VAST document itself is ordinary.
The same week, sellers are debating fill and supply-path tooling on AWS RTB Fabric. Buyers still lose hours on discrepancy tickets where seller logs show quartiles and the brand team insists nobody saw the spot. This post is about the document and player behavior HUMAN published, not about naming a botnet. If your proof of CTV delivery is a green validator on pasted XML, you are inspecting the wrong layer.
What on-screen playback would have to prove
VAST was written for players that fetch a tag, choose a MediaFile, decode video, and fire tracking pixels as the user sees the creative. Completion and quartile URLs assume the rendering surface is the surface the viewer watches. SSAI and wrapper chains complicated that contract by moving the impression pixel earlier in the chain, but the player still had to decode something for the quartiles to mean attention.
Off-screen execution breaks the implied join. The ad server still returns InLine or Wrapper XML with HTTPS trackers. The SDK still issues GETs when the internal player reaches offsets. Seller analytics can show a completed view when the only framebuffer involved was invisible. That is not a missing UniversalAdId problem. It is a player placement problem that XML alone cannot express.
HUMAN says BADBOX 3.0 also monetized spoofed server-side ad insertion on some paths without a device at all. That is a stitcher claim without a household player. The box-side modules are the mirror image: a device exists, VAST runs, and the household sees nothing unusual on the HDMI output. Both patterns bill video-shaped telemetry; only one involves MediaFile elements your validator can download.
What HUMAN documented on the wire
Satori describes ad fraud modules that mimic video ad SDKs and render in a hidden window. In one sample, a system application named com.android.systemsafe uses com.speed.adv to load three separate VAST video ad SDKs. Researchers report that the app received CTV video campaigns and played them in a transparent player outside the visible screen.
A second sample poses as GoogleAd, exposes com.google.AdActivity, and implements showAd for visible playback plus showHideWindowAd, which starts a background service and renders through a hidden view. HUMAN notes that the hidden path is complete in the APK even when a test server declined to return a configuration. The fraud stack can call the same ad servers legitimate apps call and receive similar configurations back.
Volume figures in the release measure bid requests, not confirmed views on a TV panel. HUMAN compares peak traffic to VASTFLUX and states that more than 4% of CTV traffic on its platform matched the operation at height. Those denominators are HUMAN's observed set, not the open internet. They still bound how often buyers were competing against VAST fetches that never intended living-room attention.
Who can see the gap
The publisher app on a certified Android TV device sees surface lifecycle, audio focus, and the user-visible activity. The infected AOSP box sees a system app with permission to draw outside the interactive layer. Downstream exchanges see app bundle names borrowed from well-known CTV titles and residential IPs that look like couch traffic.
A solutions engineer at a DSP can see the sample tag, the vendor documentation, and a demo player long before the live stitch hop. That person is blamed when the buyer disputes delivery even though the fraud executed after the bid response left the seat. Your preview environment almost always renders in a visible web or native player. It will fetch MediaFiles, maybe even fire start events, and never reproduce showHideWindowAd on an uncertified firmware image you do not ship to QA.
Verification companions that load inside the creative session measure the inline that played in the test harness. They do not prove the production player on a compromised box used a visible surface. OM SDK expectations assume the player reports geometry and viewability against a real container. A transparent overlay parked off-screen is outside the story those APIs were meant to tell.
<VAST version="4.2"> <Ad id="ctv-line-441"> <InLine> <AdSystem>Demand Host</AdSystem> <AdTitle>Q4 Prospect</AdTitle> <Impression><![CDATA[https://buyer.example/imp?id=441]]></Impression> <Creatives> <Creative> <Linear> <Duration>00:00:30</Duration> <MediaFiles> <MediaFile delivery="progressive" type="video/mp4" width="1920" height="1080"> <![CDATA[https://cdn.example/creative.mp4]]> </MediaFile> </MediaFiles> <TrackingEvents> <Tracking event="start"><![CDATA[https://buyer.example/track?e=start]]></Tracking> <Tracking event="complete"><![CDATA[https://buyer.example/track?e=complete]]></Tracking> </TrackingEvents> </Linear> </Creative> </Creatives> </InLine> </Ad></VAST>This shape is what a DSP sample review expects. Impression and quartile URLs use HTTPS. MediaFile carries delivery, type, and dimensions CTV players require. vastlint flags missing attributes and malformed trackers; it does not ask whether com.speed.adv drew the pixel on a surface the owner could see.
When the same XML succeeds in your lab, the handoff note says the tag is valid. The production failure happens inside com.android.systemsafe, not inside the XMLElement tree. Support then argues IVT scores or seller identity instead of player surface policy on sideloaded firmware.
Wrapper chains compound the illusion. An outer hop can return InLine trackers while an inner hop supplied the MP4. Your fetch-based review might stop at the first playable MediaFile even though the infected box never surfaced that player to HDMI.
Get VAST spec updates, platform guides, and release notes in your inbox.
<VAST version="4.2"> <Ad id="exchange-wrap-77"> <Wrapper> <AdSystem>SSP Gateway</AdSystem> <VASTAdTagURI><![CDATA[https://buyer.example/vast?deal=900]]></VASTAdTagURI> <Impression><![CDATA[https://ssp.example/wrapimp?id=77]]></Impression> <Error><![CDATA[https://ssp.example/vasterror?code=[ERRORCODE]]]></Error> <Creatives/> </Wrapper> </Ad></VAST>Error on the wrapper is present, so chain diagnostics can still fire if fetch fails. None of that reveals an off-screen final player. The buyer URL behind VASTAdTagURI may return the InLine in the first example while the device executes it invisibly.
Compare this to the BADBOX 3.0 bidstream costume described on rtblint.org: bundle names that impersonate apps the box never installed. The VAST half and the OpenRTB half fail on different evidence. A DSP must keep both reviews separate instead of collapsing them into one green checklist.
HUMAN also observed SSAI spoofing on overlapping supply paths. That failure never downloads your MediaFile. Box-side hidden VAST does. Your validator and inspector tools are relevant to the second case, but only if someone remembers validation is not proof of glass.
Sample-tag review before CTV handoff
- Fetch the live tag URL and record hop count; save the final InLine, not only the outer wrapper.
- Confirm MediaFile delivery, type, width, and height on every file the player would attempt.
- Match declared VAST version to the buyer player matrix for CTV, not only desktop QA.
- Require HTTPS on impression, error, and quartile URLs; note any http mixed-content risk.
- Compare AdServingId and creative identifiers across hops for reuse that breaks log joins.
- Document that lab playback used a visible player; note that sideloaded AOSP firmware is out of scope for tag XML alone.
- Escalate bundle and seller authorization questions to the OpenRTB review, not to VAST linter output.
- Pair seller discrepancy exports with the rtblint check on app.bundle versus sellers.json when impersonation is suspected.
What to send when delivery is disputed
Send the fetched VAST chain, the player surface policy for the claimed environment, and whether the inventory path allows uncertified AOSP devices. Send verification logs from the disputed window, not only the creative approval PDF from launch week.
vastlint checks malformed XML, wrapper depth, missing Error URIs, and tracker URL hygiene. It does not detect hidden windows or firmware backdoors. Treat a clean linter pass as necessary, not sufficient, when the buyer alleges off-screen playback.
Cross-check the seat discrepancy report against HUMAN's published indicators: concentrated traffic on a handful of box models, proxy-heavy IP profiles, and hour-aligned volume spikes. Those are bidstream signals. The VAST evidence still matters when the dispute is whether your tag could have produced quartiles without visible rendering.
Validate the tag shape before you argue playback
Paste sample XML or fetch the buyer URL, then walk wrapper hops in the inspector when the first response is still a Wrapper.
Validate VAST