Nov 16 / Latest News

ArcticRAT Malware Uses Grok to Mine Stolen Bank Texts

An Android remote access trojan called ArcticRAT has begun handing stolen text messages to a commercial AI model so its operators can find money faster, according to research Zimperium's zLabs team published on November 12, 2026. The malware sends a victim's installed-app list and harvested SMS messages to xAI's Grok, which flags banking, cryptocurrency and trading apps and picks out login codes, balance updates and notices of money moving. Zimperium said the AI feature first showed up in samples analyzed around mid-July 2026 and that the malware is still under active development.

ArcticRAT did not start out this capable. Zimperium traced it back to spyware that went after people using 11 particular crypto apps, and took its name from early package identifiers including net.iceflow.arcticlink and net.frostlink.iceforge. Over the following months the code grew into a modular tool for full device takeover. The researchers collected 78 unique samples, made up of 42 droppers and 36 payloads, along with 21 command-and-control addresses and 129 distinct commands. Builds labeled v6.3 through v6.5.2 showed the developers shipping changes steadily, and most of the server addresses Zimperium mapped were in Hong Kong, with a smaller share in Germany.

Victims were reached through malvertising, smishing and links posted on third-party forums. Those led to landing pages dressed up as Google Play, offering fake versions of Amazon, TikTok, file manager and crypto wallet apps, and some payloads were served straight from Amazon Web Services storage buckets. The dropper used a neat trick to stay quiet. It asked for permission to run as a VPN, then routed all traffic to nowhere, which Zimperium said kept Google Play Protect from reporting home while the real payload was decrypted and installed. Once that was done, the VPN was switched off.

The AI step is narrow but useful to a thief. When operators send a command named xAI_API, the server pushes an API key down to the phone, and the malware runs two prompts built into its code. One reviews the app list for financial software. The other strips promotional clutter out of the stolen texts, keeps the codes and account alerts, and estimates how well off the victim appears to be, returning the answer as structured JSON. Zimperium observed calls to a Grok model identified as grok-4.20-0309-non-reasoning. The results sit in a hidden folder on the device until a second command collects them. The model only sees data the malware has already taken. It does not touch bank accounts directly, but it spares the operators the manual sorting.

The rest of the toolkit is aimed at using those codes without the owner noticing. After a victim grants Accessibility permissions, ArcticRAT logs keystrokes into daily JSON files, records taps and screenshots, streams the screen and microphone, and opens a hidden VNC session that lets an operator move through banking apps in real time. During that work it can lay a black overlay across the display and set brightness to zero, so the phone looks switched off. A fake lock screen captures the user's PIN, pattern or password, and the malware checks the stolen value against Android itself with a built-in system command. A separate module shows a ransom screen with a progress bar that never ends, although nothing is actually encrypted.

Newer builds went further by abusing Wireless Debugging, an Android 11 feature that lets developers connect over Wi-Fi. ArcticRAT scans the local network for the pairing service and, once an operator supplies the six-digit pairing code, opens a privileged shell. From there it can silently switch on its own Accessibility access, read the lock settings, strip the lock screen entirely or set a new code that locks the real owner out of the device. No software vulnerability is involved at any stage, so there is no patch to wait for. Every step depends on the user installing an app from outside the official store and granting the permissions it requests.

The campaign also points to a shift in how criminal groups divide labor. Picking through thousands of texts for the few that matter used to take a person time, and that delay gave fraud teams a window. Pushing the work to a hosted model removes much of it. Zimperium described the change as mirroring what it had seen in RatHat, another malware family, which suggests the approach is spreading rather than being one group's experiment. Using a mainstream AI service also means the traffic can look like ordinary calls to a legitimate API.

The incident underscores how little protection a texted passcode offers once the phone that receives it is in someone else's hands. NIST's digital identity guidance has listed SMS codes as a restricted authenticator since 2017, and this case shows why. Banks and insurers should review where SMS still serves as the only second factor for high-risk actions such as new payees, address changes and large transfers, and should test whether their apps can detect overlays, screen sharing and Accessibility abuse during a session. Mobile device management policies for staff phones should block installs from unknown sources and disable Wireless Debugging. Auditors can ask whether fraud analytics treat a fast sequence of OTP use, payee setup and transfer from a familiar device as suspicious rather than trusted. Risk teams should also add outbound calls to public AI APIs from unexpected apps to their monitoring, since that pattern is now a known sign of compromise.


Make the description more conciseAdd a defender-focused sectionFormat the IOCs as a table AllPeoplePeopleFilesFilesMeetingsMeetingsEmailsEmails
Thank you!