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 at scale.

The threat was discovered by Dutch mobile security firm ThreatFabric in late October 2026 during an ongoing investigation into a separate NFC-relay banking Trojan. According to ThreatFabric’s technical report, Bazuka operates as a full malware framework. Its architecture is divided into a dropper and downloader component, a core module, and 23 supplementary modules that collectively handle malicious capabilities such as screen overlays, keylogging, two-factor authentication (2FA) and SMS interception, remote device control, and cryptocurrency address swapping.

What distinguishes Bazuka from many traditional mobile trojans is a novel overlay technique. Instead of relying purely on standard transparency or system-level tricks, the malware can capture a screenshot of the user’s home screen or lock screen and display that static image as an overlay on demand. This effectively masks malicious activities taking place quietly in the background while the user believes their device is sitting idle or locked.

Furthermore, the malware incorporates artificial intelligence directly into its command-and-control (C2) panel to help human operators identify high-value targets of interest and manage infections more efficiently. The framework supports both locally hosted AI models and those accessed via external APIs. Sample prompts recovered by researchers indicate that operators can instruct the AI model to score victims based on specific parameters, determining how attractive or lucrative a particular infected device is as a target for financial theft.

Bazuka Android Malware Uses Fake Screen Overlays to Hide Credential Theft and Device Control

The infrastructure supporting Bazuka indicates that dropper applications can be distributed to victims through at least two primary channels. Threat actors known as "traffers" can distribute the payloads by using social media advertisements or fraudulent websites to trick users into downloading malicious packages, while other distribution campaigns have leveraged the official Google Play Store.

Among the discovered dropper variants, one particularly notable sample was found masquerading as a premium, adult-oriented version of the popular social media application TikTok, utilizing the package name "com.tiktok.premium." Security telemetry has revealed at least two distinct variants of the Bazuka attack chain, which rely on stagers and downloaders to quietly fetch the core components and necessary supplementary modules required for a given campaign before launching them on the target device.

Once the core and required modules are successfully loaded, Bazuka assumes the role of a persistent keylogger and classic overlay trojan. It continuously monitors text input entered by the victim and deploys targeted overlays precisely when sensitive banking or authentication fields are accessed. All operational functionality is dynamically directed by commands issued from the remote C2 infrastructure.

The core module acts as the orchestrator of the entire framework. It maintains continuous communication with the C2 servers, loads additional dynamic execution (DEX) modules, dispatches operational commands, gathers comprehensive device telemetry, and manages application permissions alongside component updates. Mirroring tactics observed in other advanced Android trojans, Bazuka actively abuses Android Debug Bridge (ADB) shell commands and the operating system’s built-in accessibility services to request and automatically grant itself elevated permissions without requiring direct user consent.

Bazuka Android Malware Uses Fake Screen Overlays to Hide Credential Theft and Device Control

Critical components within the core include mechanisms designed to enable Wireless Debugging and complete the pairing process covertly. The core module can automatically open the system settings screen, navigate to the relevant developer menu, enable Developer options, and subsequently switch on Wireless Debugging. Once this mode is active, the malware intercepts the necessary PIN, which is then utilized to establish a secure Transport Layer Security (TLS) connection directly to the ADB daemon.

Additionally, the core component employs defense-evasion routines specifically designed to counter detection mechanisms implemented by Google Play Protect by disabling protective components through the accessibility services API. It is also equipped with a built-in self-destruct mechanism capable of wiping local storage, deleting its database entirely, and disabling the host application to erase forensic artifacts if interference or discovery is suspected.

Communication between the infected device and the C2 server is established over dual WebSocket channels to ensure persistent connectivity and command delivery. Once connected, operators can deploy any of the 23 supported modules, enabling a comprehensive suite of data harvesting and cryptocurrency theft capabilities designed to compromise financial accounts and personal information.

While the malware has not yet been definitively attributed to any known threat actor or organized cybercrime group, analysis of the operator control panel revealed user interface elements rendered in a mixed combination of English and Russian. ThreatFabric researchers noted that the integration of ADB bridging and LLM-powered management consoles represents a concerning evolution in modern mobile threat engineering.

Bazuka Android Malware Uses Fake Screen Overlays to Hide Credential Theft and Device Control

ArcticRAT Android Malware Also Integrates LLMs to Extract High-Value Data

Bazuka is not an isolated development in the deployment of artificial intelligence within mobile threats. Just prior to its disclosure, researchers from the Zimperium zLabs research team detailed another rapidly evolving and sophisticated Android malware known as ArcticRAT. This threat supports 129 distinct operational commands and incorporates LLM APIs directly into its data exfiltration pipeline, allowing operators to automatically parse and extract high-value financial intelligence from stolen SMS messages and application databases.

According to Zimperium researchers Rajat Goyal and Fernando Ortega, ArcticRAT evolved from what initially began as a targeted spyware operation focused on users of 11 specific cryptocurrency applications into an enterprise-grade, AI-powered Remote Access Trojan.

The infection chains for ArcticRAT typically rely on malicious websites meticulously designed to impersonate legitimate Google Play app listings. Unsuspecting users are frequently lured to these pages through malvertising, smishing campaigns, or links distributed across third-party forums. The malicious applications frequently masquerade as well-known brands such as Amazon and TikTok, legitimate cryptocurrency applications, or generic file manager utilities.

Upon installation, ArcticRAT exploits Android accessibility services to establish remote control over the device, automate fraudulent financial transactions, and leverage AI models to clean and categorize exfiltrated data. The malware utilizes raw TCP sockets to communicate reliably with its command-and-control infrastructure.

Bazuka Android Malware Uses Fake Screen Overlays to Hide Credential Theft and Device Control

As part of its initial reconnaissance phase, ArcticRAT builds a detailed host profile of the infected device, which dictates how the C2 server dynamically selects and deploys subsequent attack payloads. Among its defensive capabilities, the malware requests VPN service privileges to effectively null-route network communication on the victim’s device, temporarily blinding Google Play Protect telemetry during the critical staging phase.

To prevent the user from noticing or interfering with background operations, ArcticRAT employs a defense-evasion tactic referred to as a "Black Screen" overlay. When an attacker is actively navigating banking applications via Hidden VNC or auto-approving system permission prompts, the malware renders a completely black, un-dismissible view over the display and drops screen brightness to zero. To the victim, the smartphone appears completely locked, powered off, or unresponsive.

Recent iterations of ArcticRAT have also been observed abusing native Wireless Debugging capabilities to initiate privileged shell sessions on the victim’s machine. Through this shell access, the malware executes commands to toggle accessibility permissions, verify lock-screen credentials, and deploy black screen overlays to conceal malicious background activity.

Security analysts warn that the convergence of real-time Hidden VNC capabilities, raw TCP-based command channels, automated defense evasion overlays, and integrated large language models signifies a broader shift toward fully automated, high-efficiency mobile cybercrime.