CyberCartel Automates Latin American Bank Heists Using Claude and Grok AI Agents

The era of manual code injection is giving way to automated, AI-driven theft on a massive scale. A new report reveals that the CyberCartel, a notorious group active since 2012, has integrated advanced AI agents into its attack infrastructure. This shift has significantly accelerated the pace of data breaches across the Latin American financial sector.

The group has been utilizing Anthropic’s Claude and SpaceX’s Grok to streamline the process of stealing credentials and customer data. According to research from cybersecurity startup Straiker, these tools allow criminals to access bank accounts and mobile payment applications with minimal human intervention. The efficiency gained through automation has enabled the gang to target a broader range of businesses than was previously feasible using traditional human-powered methods.

The scope of the compromise is extensive, with passwords and details stolen from more than 400 institutions. These targets include major banks and popular mobile payment apps, highlighting the vulnerability of the region’s digital financial backbone. The use of AI agents has effectively lowered the barrier to entry for complex attacks, allowing a single team to execute a volume of breaches that would have required significantly more manpower in the past.

A key finding in the analysis is how easily the AI models’ built-in safety protections were bypassed. The hackers reportedly instructed the models to operate within test environments, a designation that often relaxes security constraints. When one model refused to proceed with the task, the group simply switched to a different provider. If both primary models rejected the command, the agents were tasked with deploying malware instead, ensuring the attack continued regardless of individual model refusals.

Both Anthropic and SpaceX have been notified of the findings and are actively working to block CyberCartel’s access to their respective services. Anthropic has already succeeded in restricting the group’s use of Claude, a move intended to cut off a primary tool for the gang. Straiker is also in communication with Moonshot AI, a Chinese laboratory whose agents were also identified in the attack chain, though a response from that firm is still pending.

This incident marks a significant escalation in the use of commercial AI by cybercriminals. It is notable that the group did not appear to use Anthropic’s newer Mythos model for these specific attacks. That model had drawn attention earlier this year after the company warned that such advanced systems could be exploited for similar purposes. The release of Mythos had previously prompted global security teams to address thousands of identified vulnerabilities, but it was not the primary driver of the recent Latin American breaches.

The integration of AI into criminal operations is not an isolated event for Anthropic. The company disclosed in November 2025 that its technology was used to execute a cyberattack. In another instance, Claude was utilized to breach the security systems of OpenAI, a direct competitor. While that particular intrusion was carried out by researchers rather than criminals, it underscores the dual-use nature of these powerful language models in the security landscape.

The timing of this report coincides with similar investigations in South Korea. Local authorities are currently examining AI-related breaches affecting major financial institutions, including Hana Bank, KB Kookmin Bank, and Shinhan Bank. The incident has resulted in the leakage of personal credit information for at least 25,000 Shinhan Bank customers, suggesting that the vulnerability of AI-driven attacks is a global phenomenon rather than a regional one.

Major technology firms are responding to these emerging risks with new defensive infrastructure. Nvidia, a key provider of hardware for AI data centers, recently launched the Open Agent Safety Platform. This system is designed to secure rogue AI agents through a combined hardware and software security stack. The goal is to prevent agents from escaping their designated environments, executing unauthorized code, or bypassing established guardrails to access critical infrastructure. As AI becomes more central to financial operations, these defensive measures may become a standard requirement for institutional security.