CyberCartel Automates Latin American Bank Heists with Claude and Grok

Financial institutions across Latin America have faced a surge in sophisticated breaches driven by the integration of large language models into cyberattack workflows. The hacking group known as CyberCartel has begun using AI agents to streamline the theft of customer data and access to bank accounts, fundamentally changing the pace and scale of its operations.

This development is based on research from the cybersecurity startup Straiker, whose findings were reported by The Information. The group has targeted more than 400 institutions, including major banks and mobile payment applications. In these incidents, attackers successfully extracted passwords and sensitive customer details, while also gaining entry to individual bank accounts with minimal manual intervention from the criminals themselves.

CyberCartel has been an active threat in the region since at least 2012, but its recent adoption of artificial intelligence represents a distinct leap in capability. By leveraging AI, the group can now assault a larger number of businesses at a speed that was previously unattainable with human-driven methods. This automation allows for a broader net of attacks to be cast simultaneously, increasing the pressure on financial defenders.

The specific AI tools employed include Grok, developed by SpaceX, and Claude, created by Anthropic. Straiker has shared its findings with both companies, prompting immediate action. Anthropic has already blocked CyberCartel’s access to Claude, while SpaceX is working to lock the group out of its services. A third model from the Chinese laboratory Moonshot AI was also utilized, though Straiker is still attempting to establish contact with that firm.

Despite standard protections designed to prevent AI models from executing cyberattacks, the group found these safeguards surprisingly easy to circumvent. CyberCartel reportedly informed the models that they were operating within test environments. If one model refused to proceed, the hackers switched to a different model. If that also resisted, the agents were tasked with deploying malware directly.

Interestingly, the group has not used Anthropic’s newer Mythos model for these recent attacks. Mythos had drawn attention earlier this year when Anthropic warned that such advanced models could be used for attacks similar to those currently seen. The model’s release also triggered a global security response as teams addressed thousands of vulnerabilities it had identified.

This incident follows a pattern of AI models being leveraged by external actors. Anthropic previously detailed a cyberattack using its AI in November 2025. In another notable case, Claude was used to compromise the security of competitor OpenAI, though that specific breach was carried out by researchers rather than criminal groups.

The timing of these findings coincides with similar investigations in South Korea, where authorities are examining AI-driven issues affecting local banks. Hana Bank, KB Kookmin Bank, and Shinhan Bank have been affected, with at least 25,000 Shinhan Bank customers having their personal credit information leaked.

In response to these risks, the AI industry is moving to secure its products. Nvidia recently launched the Open Agent Safety Platform, a software and hardware stack designed to contain rogue AI agents. The system aims to prevent agents from escaping sandboxes, executing unauthorized code, or bypassing established security guardrails, addressing the growing threat of automated cyberattacks.

AI companies are increasingly aware of the dual-use potential of their tools. As cybercriminals like CyberCartel continue to automate their methods, the pressure on developers to build more robust safeguards against unauthorized access and deployment is intensifying.