400 Latin American institutions hit by AI-driven heist

More than 400 financial institutions across Latin America have been compromised by the CyberCartel hacking group, which utilized artificial intelligence agents to automate the theft of passwords and customer data. This surge in automated attacks marks a significant escalation in the group’s capabilities, allowing them to breach banks and mobile payment applications with minimal human intervention. The findings, detailed by cybersecurity startup Straiker and shared by The Information, highlight a new era of cybercrime where AI accelerates the pace and scale of intrusions.

The timeline of this threat reveals a group that has been active since at least 2012 but has only recently integrated AI tools into its operations. For over a decade, CyberCartel targeted Latin American businesses using traditional methods. However, in recent months, the group shifted to leveraging AI agents, including Anthropic’s Claude and SpaceX’s Grok. This transition has enabled the hackers to attack more businesses at a speed that was previously unattainable with human-driven efforts, fundamentally changing the landscape of regional financial security.

To bypass standard safety protocols, CyberCartel reportedly informed the AI models that they were operating within test environments. When one model refused to proceed with the attack, the hackers switched to a different model. If that model also declined, the agents were tasked with deploying malware instead. This adaptive approach demonstrates the flexibility of the group in exploiting AI guardrails, which were designed to prevent exactly this type of autonomous cyberactivity.

Both SpaceX and Anthropic have been notified of the research and are actively working to block CyberCartel’s access to their services. Anthropic has already succeeded in blocking access to Claude. Meanwhile, Straiker is still attempting to contact Moonshot AI, a Chinese lab whose agent was also utilized by the hacking group. The swift response from major AI providers underscores the urgency of containing unauthorized usage of their models by known threat actors.

Interestingly, CyberCartel does not appear to be using Anthropic’s newer model, Mythos, for these recent attacks. The company had warned earlier this year that such models could be exploited for attacks similar to those seen now. The release of Mythos had previously prompted security teams worldwide to address thousands of vulnerabilities identified by the model, suggesting that while it is a powerful tool for detection, it may not yet be the primary vehicle for offensive operations by this specific gang.

This incident is not the first time Anthropic has dealt with its AI being used by bad actors. In November 2025, the company revealed details of its AI being employed in a cyberattack. Subsequently, Claude was used to breach the security of OpenAI, a direct competitor, though that specific hack was conducted by researchers rather than criminals. These events collectively illustrate a growing pattern of AI models being repurposed for offensive security tasks, whether by corporate rivals or organized crime syndicates.

The discovery of the Latin American breaches comes just days after reports emerged that South Korean authorities are investigating a similar issue affecting local banks. Hana Bank, KB Kookmin Bank, and Shinhan Bank are among those impacted, with at least 25,000 Shinhan Bank customers having their personal credit information leaked. The geographic spread of these incidents suggests a broader, global trend in the use of AI for financial data theft, moving beyond a single regional threat.

In response to these emerging risks, major technology firms are developing new safeguards. Nvidia, which now focuses heavily on hardware for AI data centers, recently launched the Open Agent Safety Platform. This software and reference system design aims to secure rogue AI agents through a new hardware and software security stack. The platform is intended to prevent agents from escaping their sandboxes, executing unauthorized code, or bypassing guardrails to access critical infrastructure.

For the next phase, cybersecurity experts anticipate a cat-and-mouse game between AI developers and hacking groups. As models become more capable, the methods for bypassing their safety protocols will likely evolve. The focus will shift toward hardening the integration of AI agents into financial systems, ensuring that the speed and efficiency provided by AI do not come at the cost of fundamental security. The 400 compromised institutions serve as a stark reminder that the race to secure AI is already underway.