Rented phishing kit uses AI to call Mexican bank victims

For 3,000 Mexican pesos a week, roughly 170 US dollars, a fraudster with no technical skill can rent a panel that puts them inside a live online banking session and lets them steer what the victim sees, screen by screen. Group-IB has published an investigation into that panel, called Balonx Sistema, which it says targets more than 20 financial institutions in Mexico and has harvested credentials and financial information from more than 1,100 users since at least October 2025. The operator's own exposed GitHub repositories gave the researchers their view of it.

A subscription, not a kit

Balonx is sold the way software is sold. Affiliates register through a Telegram bot, a SuperAdmin named balonx approves them, and subscriptions clear through the Bitso cryptocurrency exchange API. The 6,000 MXN tier adds eight sub accounts, each restricted to the banks their Admin enables. Group-IB saw about 99,000 US dollars in estimated earnings across 12 steady users. The service is advertised openly on Facebook.

The operator watches you type

What separates Balonx from an ordinary phishing page is a persistent WebSocket connection, a two way channel that stays open between the fake bank page and the criminal's console. The operator sees what the victim sees, reads what they submit, and can change their screen mid session. Group-IB documented 14 screens the operator can push on demand, among them a bank login, an SMS code prompt, an ATM PIN, card details, and one that switches on the camera to capture an ID document and a selfie.

That design exists to beat multi factor authentication. The operator replays the stolen password into the real bank site, triggers the genuine one time code, then pushes a code entry screen to the victim while the code is still valid. A "Mi Flujo Personal" feature scripts the whole sequence per bank in advance, so even an unskilled affiliate can run it. A phishing kit recently began running AI checks on the ID photos it stole, the same instinct to automate what a human used to do.

When the operator sees a target using mobile banking, they push a fake security warning urging the victim to install a bank protection app. It is an Android remote access trojan built on the commercial Spyroid framework, with its command server hidden in Base64, decoding to 196[.]251[.]84[.]11 on port 7771, over a connection with no timeout, so access is continuous.

Carolina is not a person

The finding Group-IB calls the most significant is a separate module, CallFlow, at callbalonx[.]info, which removes the human call centre altogether. A victim who answers hears a fabricated bank representative named Carolina. The voice is an ElevenLabs profile, the words come from GPT-4o-mini, and the victim's replies are transcribed live by OpenAI Whisper and fed back to the model. The calls run over a FreePBX stack at 85[.]31[.]235[.]109. Voice fraud has always been limited by how many people you can put on phones. This removes that limit, in a region where fake bank calls paired with mobile malware already work.

Call the number on your card instead

Balonx rotates domains constantly, all pointing at the same hosted database, so a takedown costs the operator a registration and nothing else. Group-IB attributes more than 350 domains to Balonx and the related "Aclaraciones Bancarias" campaigns with high confidence. Reputation based blocking will not keep up, and neither will SMS verification: banks in Mexico should assume a one time code typed into a web page is being read by an attacker in real time. Customers should treat any call or pop up asking for a code, a PIN or an app install as fraudulent, hang up, and dial the number on the back of their card.

What makes Balonx worth watching is not the engineering. It is that its operator bolted four commercial AI services onto a fraud call centre and made the human operator optional.

This briefing is provided by IntelFusions for informational and defensive purposes only. It is based on sources assessed to be reliable at the time of writing, and analytic judgments carry the confidence levels indicated. Indicators of compromise are defanged; re-arm them only in controlled environments. IntelFusions is not affiliated with the organizations named and makes no warranty as to completeness or accuracy.

Read the full analysis on IntelFusions