We Analyzed 200 Million Ad Clicks for Fraud. Here’s What We Found
Check out these 4 critical charts that show what we uncovered for this in-depth report.
NEW New report: We analyzed 200M ad clicks
The Ad Fraud Brief (Q3 2026): Real User Devices Become Fraud Proxies
- September 30, 2026
This quarter, we cover ad fraud schemes that use everyday consumer hardware as fraud vehicles, leaning on the users’ residential IPs and clean reputation to hide their activity.
Here are the biggest ad fraud schemes from Q3 2026 and what they mean for advertisers.
1. Fuyao Enterprise – Ad fraud scheme on Android TV boxes
Scale: $150,000+ estimated daily ad revenue generated by operators.
The Fuyao scheme leveraged cheap Android TV boxes to power an entire ad fraud scheme, estimated to generate $150k daily at its peak. The TV boxes were preloaded with malware from the factory and then sold to unsuspecting users.
Once powered on, the boxes switch between two modes of operation: ad fraud or residential proxy. The former is triggered any time the box is connected to HDMI and displaying on a TV. Here, it runs as a residential proxy, reselling the home owner’s internet connection.
When not connected via HDMI, the TV box clicks on ads from the operators’ own network of over 144 AI-generated sites. But not before spoofing itself into a premium phone model (like Samsung, Xiaomi, and others) to command higher ad rates.
The Fuyao Enterprise is a clear example of why ad fraud keeps evading platform filters. The TV boxes are running on real residential IPs with clean records and normal user activity, so they interact with ads without raising suspicions. Anyone renting the IP will get the same advantage: a reputation clean enough to interact with ads without triggering detection.
2. StegoAd – Malicious Microsoft Edge extension
Scale: 2.6 million installs across 119 fake browser extensions, undetected for two years
The StegoAd scheme used normal-looking Microsoft Edge extensions like ad blockers, VPNs, and video downloaders that performed exactly as advertised while dormant.
But when active, the extensions removed legitimate ads from Google AdSense and Amazon ad slots and replaced them with the operator’s own. Interestingly, the extensions only triggered this for a maximum of 6 ad placements to avoid drawing attention.
Researchers found that the Extensions in the StegoAd operation also hijacked affiliate commissions and replaced cookies with those of bad actors. Those commissions would have gone to referring sites across more than 20 Amazon country stores, eBay, AliExpress, Taobao, and JD.com.
The scheme ran for over 2 years without drawing suspicion because the campaign’s tracking infrastructure was managed through seven separate Google Analytics accounts. This story reveals how deeply a fraudulent scheme can be embedded inside trusted platforms.
The takeaway for advertisers is not to completely outsource vigilance against ad fraud. Bad actors often operate on the same trusted platforms you rely on!
3. JarService – Ad scheme on Android car screens
The JarService is the first large-scale ad fraud scheme that targeted car Android-based infotainment systems. Unlike schemes like Fuyao, where devices come preloaded with malware, JarService hijacks the target devices’ update cycle. So instead of receiving legitimate updates, the JarService payload is delivered.
Once installed, operators are able to execute several commands on the devices, including one titled “WebView” that generates fraudulent ad activity invisible to the car’s occupants and a second that turns the head unit into a reverse proxy node. Kasperky links this infrastructure to the MoYu Group, the actors behind the BADBOX botnet.
Read more about BADBOX and other massive ad fraud scams.
The mechanics of JarService mean that advertiser inventory can be clicked by a device on a completely legitimate IP address, but secretly under the control of a bad actor. Advertisers pay for engagement the real user never triggered, while the operators pocket the payout.
4. Papyrus Virus – Reading Apps Clicking on Ads
Scale: Nearly $1 million per month in monetization and 800+ domains associated with the scheme
Papyrus runs through a cluster of novel-reading apps on Android. At face value, they deliver the expected value, allowing users to read fictional novels. But underneath, the apps launch hidden browsers that load web pages and click on ads.
This part of the scheme isn’t new. What’s novel is how the scheme mimics human browsing patterns. The hidden browsers follow scripted interaction patterns like click points, scroll ranges, and timing between actions, with each action point selected at random so no two interactions are alike.
The automation also autoclicks consent and cookie banners on the pages as they load, further legitimizing the browser action.
With these tactics, Papyrus outperformed real traffic on the metrics that matter to advertisers: click success rates roughly 25 times higher than legitimate traffic, cost-per-thousand-impression rates around 4 times higher, and attention scores about 13% higher.
Advertisers viewing these metrics on a dashboard see a highly successful campaign, but it’s actually a budget black hole.
5. HuggingFace – AI agents powering fraud
Scale: 7,851% spike in AI agent traffic YoY in 2025
Last, but not least, we can’t help but mention the HuggingFace hack that went incredibly viral across global news media.
In July 2026, OpenAI’s AI agents broke out of a sandboxed security test and into Hugging Face’s production systems. During their movement through Hugging Face’s systems, they harvested credentials and used them to access deeper infrastructure. It’s been widely reported, with OpenAI publishing its own incident report.
The bigger concern for advertisers is that the agents’ pattern of behavior – evade detection, act autonomously, and use whatever access you can find to keep going – is eerily similar to how ad fraud bots operate.
Researchers tracking AI traffic found the behavioral gap between benign and malicious automation narrowed to about half a percentage point in 2025, which means there’s very little way to tell the two apart. An AI agent clicking on an ad can present with a real browser fingerprint, residential IP, and engagement patterns that look human and will slip through most filters.
The Hugging Face incident also points to another critical lesson in ad fraud detection: the answer lies in sophisticated behavioral monitoring, similar in principle to what flagged the unusual activity of OpenAI’s agents. Integrated with other checks, it remains a cornerstone for differentiating legitimate from invalid activity.
Q3 Ad fraud takeaways
The basic mechanics of ad fraud haven’t changed. Every scheme we covered this quarter still comes down to faking clicks to collect a payout that was never earned. However, we are seeing greater speed and sophistication, which lets bad actors infect a greater number of devices and stay undetected for much longer.
The exposure for advertisers has never been higher. And in the case of many schemes we covered, most of the traffic looks legitimate to surface-level scrutiny and can even outperform real campaigns on key metrics.
Dedicated click fraud protection remains one of the most reliable ways to protect your ad spend from schemes like these. Try Fraud Blocker free for 7 days and see how much invalid traffic is hiding in your campaigns.
Facebook
X
LinkedIn
Email
ABOUT THE AUTHOR
Matthew is the resident content marketing expert at Fraud Blocker with several years of experience writing about ad fraud. When he’s not producing killer content, you can find him working out or walking his dogs.
Matthew is the resident content marketing expert at Fraud Blocker with several years of experience writing about ad fraud.
More from Fraud Blocker
How Much Does Google Ads Cost? (Hidden Costs for Advertisers)
Average Google Ads CPC is $5.26, but our research shows advertisers are actually paying $5.62 due to invalid activity. Here's…
What is Click Fraud?
If your ads aren’t delivering results, click fraud could be the culprit. Fraudulent traffic is draining ad budgets on Google,…