Balonx and BREEZE COMET: Customer and Institutional Fraud

Banking fraud may target the person authorising a transaction or the institutional access used to execute it. Reports on Balonx Sistema and BREEZE COMET help examine that distinction. Their common financial theme does not make them one campaign.

This article synthesises third-party reports and then offers an editorial interpretation. Provider figures remain separate: they are neither added together nor presented as a measurement of all fraud in Mexico and Brazil.

The customer layer: Balonx Sistema

Group-IB described a phishing-as-a-service platform targeting more than 20 Mexican financial institutions in its 19 August 2026 report. Weekly subscriptions cost 3,000 MXN for an individual plan covering up to two devices, or 6,000 MXN for an office plan with up to eight subaccounts. The report describes automated payment processing through the Bitso API.

The infrastructure count needs qualification: Group-IB associates more than 350 domains with the “Aclaraciones Banciaras” and Balonx campaigns and dates the earliest recorded domain to 2019. That does not date every feature of the present platform to 2019 or establish that every domain belongs exclusively to Balonx.

The reported mechanism maintains a WebSocket connection so an operator can present any of 14 deceptive screens, including login, OTP, PIN and card prompts. The risk lies not just in the form’s appearance but in using a code while it is still valid. The PROTECCION_BANCARIA screen also offers an Android trojan based on Spyroid.

The report presents CallFlow as a voice-fraud component using GPT-4o-mini for dialogue, ElevenLabs for voice and Whisper for transcription, with 30 concurrent SIP channels configured by default. Those are reported capabilities, not evidence that every operation involved 30 observed fraudulent calls.

As a separate reference, Kaspersky documented JanelaRAT on 13 April 2026 and reported 14,739 attacks in Brazil and 11,695 in Mexico during 2025 in its own telemetry. Its discussion of live banking-session hijacking supports a comparison of mechanisms, not attribution of JanelaRAT to Balonx.

The institutional layer: BREEZE COMET

Google Threat Intelligence / Mandiant published its analysis on 1 September 2026, describing an actor tracked since 2024. The focus is organisations authorised to use banking applications, APIs and systems such as Pix, STR and Boleto. The report describes compromised privileged accounts and COBALTSPIN, a tunnel written in Rust.

Mandiant places two waves of hundreds of fraudulent transactions within 24 to 48 hours of access to core financial applications, based on a client report and an external forensic report. This is the timeframe of the described episodes, not a universal duration for financial intrusions.

The provider also describes language-model use to accelerate reconnaissance, credential validation and deployment, and overlaps with activity reported as Plump Spider and SHADOW-AETHER-064. These are Mandiant’s associations, not a link to Balonx established by this article.

My reading: controls for different attack surfaces

The editorial conclusion is that protecting customer confirmations and protecting privileged access are complementary responsibilities. The first concerns who controls the authentication interaction; the second concerns who can operate payment applications and how that access is monitored.

AI appears in different parts of the process: conversation and voice in CallFlow, technical tasks in BREEZE COMET. The reports do not establish that AI alone makes the attacks effective or that both actors form a single AI-enabled operation.

For context, consult country coverage and evidence gaps in the LATAM dataset. Documentary coverage does not measure prevalence or prove shared infrastructure.

This comparison raises two review questions: which confirmations could an attacker relay while interacting with a customer? And which institutional privileges could enable transactions before a fraud team detected the pattern? These are assessment questions, not a certification of controls or a complete prescription for every institution.