Cybersecurity

RatHat: An Android Malware Strain Using AI to Control Devices

Zimperium researchers discovered RatHat, a malware strain that employs an AI-powered engine to navigate the Android interface and remotely control the device. The malware combines the theft of banking and cryptocurrency application data with the interception of messages and one-time codes, as well as multiple mechanisms for persistence and resistance to removal.

2026-09-17
4 min read
445 views
certi.news Editorial Team
RatHat: An Android Malware Strain Using AI to Control Devices

Researchers from Zimperium’s zLabs labs discovered a new Android malware strain called RatHat, which uses an AI-powered subsystem to help its operators navigate infected devices and execute remote commands. According to the researchers, the malware spreads through malvertising, text messages, and phishing websites promoting APK files outside the Google Play Store.

The researchers believe RatHat may be linked to threat actors from China after finding prompts directed at large language models and written in Chinese. However, this observation does not represent definitive attribution of the campaign or its operators, but rather evidence within the malware analysis.

Deeper Control Through Android Permissions

Like other Android malware families, RatHat relies on the abuse of Accessibility Service permissions to perform highly privileged actions. More importantly, it enables Developer Options and wireless debugging, allowing it to establish a local execution context with ADB shell-level privileges without requiring an external computer, in a technique resembling mechanisms seen in ToxicPanda and RedHook malware.

After obtaining ADB access, the malware installs a Go-based client inside the liblocal-service.so file. This client executes commands with ADB shell privileges, bypasses battery restrictions, and manages persistence mechanisms on the device. It can also restore the malware if it is removed or stopped, while RatHat can in turn reinstall the client, creating a mutually redundant backup channel for persistence.

A second client, libmedia_codec.so, operates as an FRP reverse proxy client to create a persistent tunnel connecting the device to the attacker.

What Does AI Add?

RatHat serializes Android’s live Accessibility tree in XML format, then sends it to an AI assistant that Zimperium’s report did not name. The system is asked to identify the center coordinates of a specific interface element, extract the text actually displayed, and return navigation instructions such as SCROLL_DOWN.

In practice, this approach allows the malware to handle changing interfaces more flexibly than scenarios relying entirely on fixed text, while also reducing the need for direct and continuous operator interaction. Zimperium says this makes control more adaptive and could make it harder to detect than traditional automation based on predefined actions.

Data Theft and Resistance to Removal

RatHat displays HTML overlays over banking and cryptocurrency applications to lure victims into entering account details. It also intercepts text messages and notifications, including one-time passcodes, records text-change events, extracts URLs from the browser bar, and captures PINs, passwords, and unlock patterns.

The malware intercepts the uninstall confirmation screen to cancel the process, then displays a fake overlay mimicking Google Play and presents a false error message. In addition, it uses several anti-analysis techniques, such as tampering with the APK container, an oversized Android manifest file measuring 61 megabytes, and fake or invalid DEX instructions intended to confuse or disable analysis tools.

What Does This Mean for Users?

The use of AI alone does not prove that the malware bypasses all security measures, but it changes the practical weakness: control does not rely solely on fixed paths that can be predicted. Therefore, application installation sources and Accessibility permissions remain critical factors. Zimperium recommends not installing APK files from outside Google Play unless the publisher is explicitly trusted, avoiding granting Accessibility permissions to unnecessary applications, and conducting periodic scans with Play Protect.

News source
BleepingComputer
Open original source ↗
c
Author

certi.news Editorial Team

What you need to know

Zimperium researchers discovered RatHat, an Android malware that uses an AI-powered assistant to navigate the device interface and execute remote commands. It steals banking data and temporary codes, and exploits accessibility permissions and wireless debugging to maintain persistence.

  • RatHat spreads through malicious advertisements, text messages, and phishing websites promoting APK files outside Google Play.
  • The malware uses Android’s XML-formatted UI tree to identify screen elements, extract text, and issue navigation instructions.
  • It abuses accessibility service permissions and enables developer options and wireless debugging to obtain an ADB shell context.
  • It steals data from banking and cryptocurrency apps and intercepts messages, notifications, and one-time passcodes.
  • It uses persistence and anti-removal mechanisms, including reinstalling malware components and intercepting the uninstall screen.
  • Zimperium recommends avoiding untrusted APK files, not granting accessibility permissions unnecessarily, and using Play Protect regularly.

FAQ

How does RatHat spread?

According to the researchers, it spreads through malicious advertisements, text messages, and phishing websites promoting APK files outside the Google Play Store.

How does RatHat use AI?

It sends Android’s live accessibility tree in XML format to an AI assistant to identify interface-element coordinates, extract text, and return instructions such as scrolling down.

What data does RatHat target?

It targets banking and cryptocurrency account data, messages and temporary codes, PINs, passwords, and unlock patterns.

Has RatHat been definitively attributed to Chinese actors?

No. The researchers found Chinese-language prompts directed at large language models, but clarified that this is analytical evidence rather than a definitive attribution.

In the same category

You may also like

View all news