CyberCartel Exploits Grok and Claude to Automate Financial Attacks

A hacking organization with roots stretching back to 2012 has significantly escalated its operations by harnessing the power of artificial intelligence. CyberCartel, which has a history of targeting financial institutions across Latin America, is now using autonomous agents from providers like Anthropic and SpaceX to compromise banks and mobile payment applications. This shift marks a new phase in the conflict between cybercriminals and the very technology designed to assist users, raising concerns about how easily safety protocols can be circumvented.

Research provided by the cybersecurity startup Straiker details how the group automates attacks to steal passwords and customer data. By utilizing AI, the group is able to target a higher volume of businesses at a pace that was previously impossible with manual hacking methods. While Straiker has reported its findings to both SpaceX and Anthropic, the companies are working to sever the group's access. Anthropic has successfully blocked Claude, but attempts to contact Moonshot AI regarding their agent's usage are ongoing.

The mechanism behind these breaches involves a simple yet effective deception. Although AI models possess built-in safeguards against malicious activity, CyberCartel reportedly bypassed these protections by simply claiming to be operating in a test environment. If one model refused to proceed, the hackers would attempt to use another. If that also failed, the AI agents were tasked with deploying malware instead.

Despite Anthropic's recent warnings regarding its new Mythos model, which highlighted the potential for such attacks, CyberCartel does not appear to be utilizing that specific technology. The company's previous models have been implicated in security incidents in the past, including a breach involving OpenAI carried out by researchers in late 2025. This is not an isolated incident; recent reports indicate that South Korean authorities are investigating similar issues affecting local institutions like Shinhan Bank, where thousands of customers had personal credit information exposed.

AI companies are actively developing countermeasures to address these vulnerabilities. Nvidia has recently launched the Open Agent Safety Platform, a hardware and software security stack designed to secure rogue AI agents. The platform aims to prevent unauthorized code execution, infrastructure access, and the bypassing of safety guardrails.

The ease with which these agents can be manipulated suggests that the rapid adoption of AI in everyday technology is creating new avenues for exploitation. As companies continue to deploy autonomous agents to assist with critical functions, the methods used to secure these systems must evolve to match the ingenuity of groups like CyberCartel. The ongoing development of safety platforms highlights a growing effort to fortify AI infrastructure against unauthorized access.