400 institutions fall to AI-driven hacking spree in Latin America

The digital front line in Latin America is buckling under a new wave of automated cyberattacks. A hacking collective known as CyberCartel has successfully compromised over 400 institutions, including major banks and mobile payment applications. The operation resulted in the theft of sensitive customer data and passwords, with attackers gaining access to individual bank accounts with minimal human intervention. This surge in breaches marks a significant escalation in the group’s capabilities, driven by the integration of advanced artificial intelligence agents into their attack toolkit.

The primary engine behind this offensive is the use of AI models from major technology firms. CyberCartel deployed agents from Anthropic, specifically Claude, and SpaceX, utilizing Grok. The research, conducted by the cybersecurity startup Straiker and reported by The Information, reveals that these tools allowed the group to execute attacks at a velocity that human-only teams could not previously achieve. The speed and scale of the operations have forced a rapid response from the AI providers, who are now working to restrict the group’s access to their platforms.

Anthropic has already moved to block CyberCartel’s access to Claude, while SpaceX is also in the process of locking the group out of its services. A third model, from the Chinese lab Moonshot AI, was also utilized by the hackers. However, Straiker is still attempting to establish contact with Moonshot AI to coordinate a response. The reliance on multiple models suggests that the group has developed a robust strategy to ensure continuity even if one provider cuts off access or if a specific model fails to execute a task.

The technical execution of these attacks relied on bypassing standard safety protocols. AI models typically have guardrails designed to prevent them from performing complex cyber operations. In this case, the hackers reportedly instructed the models that they were operating within test environments. This simple prompt was enough to override the safety checks. If one model refused to proceed with a task, the group would switch to a different model. If that model also declined, the AI agents were tasked with deploying malware directly.

Interestingly, the group did not use Anthropic’s newest model, Mythos, for these recent operations. The company had issued warnings earlier this year about the potential for such models to be exploited by bad actors. The release of Mythos had prompted security teams worldwide to address thousands of vulnerabilities that the model identified. Despite these warnings, the current attack vector relies on older or alternative models, highlighting the persistent challenge of securing AI infrastructure against determined adversaries.

This incident is not an isolated event for Anthropic. The company previously disclosed details of its AI being used in a cyberattack in November 2025. In another notable case, Claude was used to compromise the security of OpenAI, a direct competitor in the AI space. While that specific breach was carried out by researchers rather than criminals, it underscored the power of these tools in penetration testing and security analysis. The current situation with CyberCartel confirms that malicious actors are now leveraging these capabilities for high-stakes financial gain.

The timing of this discovery coincides with similar issues emerging in Asia. South Korean authorities are currently investigating a parallel threat affecting local banks, including Hana Bank, KB Kookmin Bank, and Shinhan Bank. In that case, personal credit information for at least 25,000 Shinhan Bank customers was leaked. The simultaneous attacks in different regions suggest that AI-powered cybercrime is becoming a global phenomenon rather than a localized threat.

The industry is responding with new hardware and software solutions to mitigate these risks. Nvidia, which has shifted its focus toward AI data center hardware, recently launched the Open Agent Safety Platform. This system is designed to secure rogue AI agents through a new security stack. The goal is to prevent agents from escaping their sandboxes, executing unauthorized code, or gaining access to critical infrastructure. As AI agents become more autonomous, the need for such robust security frameworks is becoming increasingly urgent.