Cybersecurity researchers have uncovered details of a previously unknown and sophisticated Android malware framework named Bazuka, which introduces large language model (LLM) integration to streamline victim management and control operations at scale. The discovery highlights a growing and concerning trend among threat actors who are increasingly turning to artificial intelligence to automate, optimize, and scale mobile banking fraud and data exfiltration.
According to technical findings published by Dutch mobile security firm ThreatFabric, the threat was first identified in late October 2026. Security analysts stumbled upon the framework while conducting an ongoing investigation into a separate Near Field Communication (NFC) relay banking Trojan that has been active in the threat landscape. Rather than a simple, single-purpose application, researchers quickly realized that Bazuka is an extensive, full-fledged malware ecosystem built to handle multiple phases of an attack chain.
ThreatFabric’s technical report notes that Bazuka comprises a sophisticated dropper and downloader component, a core orchestration module, and a staggering 23 supplementary modules. These individual modules are designed to cover a wide array of malicious activities, ranging from screen overlays and keylogging to two-factor authentication (2FA) interception, SMS monitoring, remote device control, and malicious cryptocurrency-address swapping intended to divert digital funds.
One of the most notable aspects of the malware that distinguishes it from traditional Android trojans is a novel overlay technique. Instead of relying purely on standard drawing methods, the Trojan can capture a high-resolution screenshot of the victim’s home screen or lock screen. It can then display this captured image dynamically as an overlay on demand, effectively masking any malicious activities, unauthorized financial transactions, or data collection taking place silently in the background while the user believes their device is idle or secured.

Perhaps most alarming to security professionals is the incorporation of artificial intelligence directly into the malware’s command-and-control (C2) operational panel. This AI integration assists human operators in identifying high-value targets, prioritizing infected devices, and managing ongoing infections with minimal manual overhead. The framework’s control panel reportedly supports both locally hosted machine learning models and models accessed via external APIs, giving cybercriminals flexible infrastructure options.
Sample prompt configurations recovered by researchers indicate that operators can instruct the integrated AI model to analyze and score victims based on specific behavioral or financial criteria. This automated scoring system allows attackers to determine which compromised devices represent the most lucrative targets for fraudulent bank transfers or deep surveillance.
The distribution network for Bazuka appears multifaceted, utilizing multiple vectors to maximize reach. Threat actors rely heavily on "traffers"—affiliates who drive traffic via social media advertisements, deceptive websites, or third-party forums—to lure victims into downloading initial dropper applications. In addition to these traditional social engineering tactics, investigators have observed malicious payloads finding their way into official distribution channels like the Google Play Store.
Among the identified distribution chains, one dropper app was found disguised as a premium, adult-oriented variant of the popular social media platform TikTok, utilizing the malicious package name "com.tiktok.premium." Regardless of the initial entry point, the attack typically follows one of several distinct chains designed to download, stage, and execute the core malicious framework without raising immediate suspicion from the host operating system.
Once a victim executes the dropper, a stager downloads the necessary core components and supplementary modules tailored for the specific campaign. ThreatFabric explains that once these essential pieces are loaded, Bazuka operates simultaneously as a keylogger and an advanced overlay trojan, carefully monitoring user keystrokes and timing its deceptive overlays to intercept credentials as victims log into banking applications or cryptocurrency wallets. All operational commands are relayed dynamically from centralized C2 infrastructure.

The core module acts as the primary orchestrator of the infection. It maintains constant communication with the C2 servers, loads additional Dalvik Executable (DEX) modules dynamically, dispatches operational commands, gathers comprehensive device telemetry, manages required system permissions, and handles component updates. Like many contemporary Android banking trojans, Bazuka aggressively abuses Android Debug Bridge (ADB) shell commands and the operating system’s built-in accessibility services to request and automatically grant itself elevated privileges without requiring explicit user consent.
Critical functions within the core module include sophisticated mechanisms for enabling Wireless Debugging and successfully completing the pairing process on targeted devices. The core can programmatically navigate through system settings menus, activate Developer options, and subsequently enable Wireless Debugging. By intercepting the system PIN during this automated process, the malware can establish a secure Transport Layer Security (TLS) connection directly to the device’s ADB daemon, cementing deep persistence and control.
Furthermore, the core component incorporates defensive capabilities specifically engineered to evade detection by security monitoring tools like Google Play Protect. By leveraging the accessibility services API, the malware can disable competing protection components or self-destruct if it detects an analysis environment. This self-destruct mechanism is capable of purging local storage, wiping associated databases, and disabling the application entirely to destroy forensic artifacts.
Communication between the infected device and the C2 server is established over robust WebSocket channels, ensuring continuous and reliable data transmission. Beneath this framework, the 23 supported modules provide attackers with an extensive toolkit for comprehensive data theft, credential harvesting, and asset looting.
Although the malware has not yet been definitively attributed to a specific, known threat actor group, analysis of the operator control panel revealed user interface elements written in a mixed combination of English and Russian. Security researchers emphasize that the integration of ADB bridging and AI-driven victim management marks a dangerous evolution in mobile malware engineering.

ArcticRAT Android Malware Uses AI to Extract High-Value Data
Bazuka is not an isolated incident in the rapidly shifting landscape of AI-enhanced mobile threats. Just weeks prior to its discovery, researchers from the Zimperium zLabs research team detailed another sophisticated Android threat known as ArcticRAT. This rapidly evolving Remote Access Trojan supports a massive arsenal of 129 distinct commands and integrates large language model APIs directly into its data exfiltration pipeline to intelligently parse, clean, and extract high-value financial intelligence from stolen SMS messages and application data.
Zimperium researchers noted that what originally began as a narrowly targeted spyware operation focused primarily on users of 11 specific cryptocurrency applications has quickly transformed into an enterprise-grade, AI-powered cybercrime tool.
The infection chain for ArcticRAT heavily utilizes deceptive websites mimicking authentic Google Play app listings. Unsuspecting users are frequently lured to these malicious landing pages through malvertising campaigns, smishing text messages, or links shared across third-party forums and messaging channels. To maintain the illusion of legitimacy, the malicious applications frequently masquerade as well-known global brands, including Amazon and TikTok, as well as mainstream cryptocurrency wallets and generic file management utilities.
Upon successful installation, ArcticRAT exploits Android accessibility services to establish remote administrative control, automate fraudulent user interactions, and harness artificial intelligence to process harvested information in real time. The malware utilizes raw TCP sockets to maintain communication with its command-and-control infrastructure.
As an initial reconnaissance step, the malware profiles the host device, transmitting detailed telemetry back to the C2 server so that operators can dynamically select and deploy the most effective attack payloads.

To maintain stealth during the initial staging phase, ArcticRAT requests VPN service privileges, which it uses to null-route network communications on the victim’s device. This effectively blinds Google Play Protect telemetry and prevents the operating system from reporting suspicious activity or downloading security updates.
When attackers need to perform sensitive actions—such as navigating through banking applications via Hidden VNC or auto-approving critical system permissions—the malware deploys a defense evasion tactic referred to as a "Black Screen" overlay. During these operations, the trojan draws a completely black, un-dismissible view across the display and drops the screen brightness to zero. To the victim, the smartphone appears completely locked, powered off, or unresponsive, masking the fraudulent transactions occurring in the background.
Recent iterations of ArcticRAT have also been observed abusing native Wireless Debugging capabilities to initiate privileged shell sessions on compromised devices. This allows operators to execute deep system commands, manipulate accessibility permissions, verify lock screen credentials, and maintain the concealing black screen overlay.
Security analysts warn that the wholesale abandonment of rigid, hardcoded attack scripts in favor of real-time Hidden VNC, raw TCP command-and-control channels, and aggressive AI-driven data processing represents a significant leap forward in the capabilities of mobile cyber threats, leading to more thorough and devastating device takeovers.
