Security researchers have identified a new category of Android malware that uses artificial intelligence to carry out advertising fraud without the user’s knowledge. The malicious software belongs to a recently observed group of click-fraud trojans that rely on machine learning rather than traditional scripted techniques.
Instead of using hard-coded JavaScript instructions to interact with web pages, this malware analyzes advertisements visually. By examining what appears on the screen, it can decide where to tap, closely imitating normal user behavior. This approach allows the malware to function even when ads frequently change layout, include video content, or are embedded inside iframes, which often disrupt older click-fraud methods.
The threat actors behind the operation are using TensorFlow.js, an open-source machine learning library developed by Google. The framework allows trained AI models to run inside web browsers or server environments through JavaScript. In this case, the models are loaded remotely and used to process screenshots taken from an embedded browser.
Researchers from mobile security firm Dr.Web reported that the malware has been distributed through GetApps, Xiaomi’s official application store. The infected apps are mainly games. In several cases, the applications were initially uploaded without harmful functionality and later received malicious components through software updates.
Once active, the malware can run in what researchers describe as a “phantom” mode. In this mode, it opens a hidden browser based on Android’s WebView component. This browser loads a webpage containing advertisements and a JavaScript file designed to automate interactions. The browser operates on a virtual screen that is not visible to the device owner. Screenshots of this screen are repeatedly captured and sent to the AI model, which identifies relevant ad elements and triggers taps that appear legitimate.
A second operational mode, referred to as “signalling,” gives attackers direct control. Using WebRTC technology, the malware streams a live video feed of the hidden browser to the threat actor. This allows them to perform actions such as tapping, scrolling, or entering text in real time.
Dr.Web identified multiple infected games hosted on Xiaomi’s platform, including titles with tens of thousands of downloads. Beyond official app stores, the malware has also been found in modified versions of popular streaming applications distributed through third-party APK websites, Telegram channels, and a Discord server with a large subscriber base. Many of these apps function as expected, which reduces user suspicion.
Although this activity does not directly target personal data, it still affects users through increased battery drain, higher mobile data usage, and faster device wear. For cybercriminals, however, covert ad fraud remains a profitable operation.
Security experts advise Android users to avoid downloading apps from unofficial sources and to be cautious of altered versions of well-known apps that promise free access to paid features.
