Straiker: CyberCartel Uses Claude and Grok to Attack LatAm Institutions

CyberCartel Deploys Autonomous Agents Against Latin American Banks

As enterprise adoption scales, cybersecurity startup Straiker has revealed that the long-standing hacking collective CyberCartel is using artificial intelligence agents—including xAI’s Grok and Anthropic’s Claude—to automate complex cyberattacks against more than 400 financial institutions across Latin America. According to reporting from The Information, the threat group leverages automated systems to breach mobile payment apps, steal passwords, and compromise user bank accounts with minimal manual input.

The Tech TL;DR:

• The Threat: CyberCartel uses multi-agent frameworks to automate credential harvesting and target reconnaissance across hundreds of Latin American financial platforms.
• The Stack: The attack framework, identified as CARLIY0S, orchestrates models from Anthropic, xAI (Grok), and Moonshot AI via pentesting tools and proxy networks.
• The Defense: Anthropic and SpaceX are actively blocking unauthorized access, while Nvidia has introduced hardware-level security stacks to contain rogue autonomous agents.

The Architecture of the CARLIY0S Attack Framework

Operating under the banner of the Tr4nsHack Team and Hexagon Security Labs Mexico, CyberCartel utilizes an automated framework designated as CARLIY0S to link AI models with software routines and residential proxy networks in order to reduce the time spent on manual tasks. Rather than operating as isolated chatbots, the AI models are assigned distinct functional roles within a pipeline. Specific Claude instances analyze target infrastructure and propose attack vectors, while secondary models review these proposals before final execution. Concurrently, Grok models assist in aggregating stolen credentials into a centralized database cataloged as FinancialSoft, which researchers found to contain thousands of sensitive access records.

# Conceptual Multi-Agent Request Flow for CARLIY0S Framework
import requests
import json

def dispatch_recon_agent(target_endpoint, payload_config):
    headers = {"Authorization": "Bearer SECURE_PROXY_TOKEN"}
    payload = {
        "agent_role": "target_analyzer",
        "target": target_endpoint,
        "mode": "simulated_environment"
    }
    response = requests.post("https://api.internal-orchestrator.local/v1/execute", headers=headers, json=payload)
    return response.json()

Bypassing Guardrails via Test Environment Spoofing

Standard enterprise safety mitigations built into foundational models have proven straightforward to bypass during hostile operations. The threat actors routinely frame their malicious prompts as authorized security testing, academic research, or software deployments within isolated test environments. When one commercial model declines a request due to built-in safety guardrails, the orchestrator automatically routes the query to an alternative lab’s model. Moonshot AI’s Kimi models have reportedly served as fallback agents when primary APIs reject unauthorized prompts. This operational resilience allows small groups of operators to scale attacks across hundreds of banking and payment infrastructure endpoints simultaneously.

Infrastructure Providers and Hardware Makers Deploy Security Defenses

Following disclosures by Straiker, infrastructure providers have initiated defensive lockdowns. Anthropic has successfully blocked unauthorized access vectors tied to Claude within the malicious framework, and ongoing technical coordination continues with SpaceX regarding Grok integrations. To address the broader systems security risk of autonomous agent breakouts, hardware manufacturers are reshaping enterprise infrastructure stacks. Nvidia recently deployed the Open Agent Safety Platform, a hardware and software reference architecture designed to prevent autonomous routines from executing unverified system code, bypassing container boundaries, or compromising critical infrastructure credentials.

Disclaimer: The technical analyses and security protocols detailed in this article are for informational purposes only. Always consult with certified IT and cybersecurity professionals before altering enterprise networks or handling sensitive data.

More on this story: PoeLLM Malware Infects Over 3,400 Servers, Lumen Researchers Say · FortiGuard Labs: AI tools shorten flaw exploitation in Latin America