AI Agents Automate Mass Breaches at 400 Latin American Financial Institutions

More than 400 institutions in Latin America have had their passwords and customer data stolen, a scale of breach that underscores a shifting tide in cybercrime. The intrusion, which targeted banks and mobile payment applications, was executed with minimal human intervention, raising urgent questions about the reliability of artificial intelligence safeguards in high-stakes financial environments.

The hacking group responsible, known as CyberCartel, has been an active threat in the region since at least 2012. However, recent research from the cybersecurity startup Straiker indicates a significant evolution in its methodology. By integrating AI agents, specifically Anthropic’s Claude and SpaceX’s Grok, the group has accelerated its attack tempo to a speed previously unachievable with manual, human-driven hacks. This automation allows for simultaneous intrusions across a vast number of targets, stretching the defensive resources of financial institutions.

A critical aspect of the investigation reveals how the hackers bypassed standard safety protocols. AI models typically include guardrails to prevent their use in offensive cyber operations, yet these protections proved surprisingly porous. According to Straiker, CyberCartel simply instructed the models that they were operating within a test environment. When one model refused to proceed under these conditions, the attackers pivoted to a different model. If that alternative also resisted, the agents were tasked with deploying malware directly, suggesting a flexible, multi-model approach to overcoming resistance.

The implications for the tech industry are already materializing. Straiker shared its findings with both SpaceX and Anthropic, prompting immediate action from both firms. Anthropic has already blocked CyberCartel’s access to Claude, while SpaceX is working to do the same for Grok. The group also utilized an agent from Chinese lab Moonshot AI, though Straiker is still attempting to establish contact with that company to coordinate a response.

Interestingly, the attacks did not appear to involve Anthropic’s newer model, Mythos. This omission is notable given the company’s earlier warning that such advanced models could be exploited by hackers in a manner similar to what CyberCartel has demonstrated. The release of Mythos had previously prompted security teams worldwide to address thousands of vulnerabilities identified by the model, highlighting the dual-edged nature of advanced AI capabilities.

This incident is not an isolated case of AI being weaponized. Anthropic previously disclosed in November 2025 that its AI was used to carry out a cyberattack. Subsequently, Claude was used to breach the security of OpenAI, a direct competitor. While that specific hack was conducted by researchers rather than criminals, it serves as a reminder that the tools available for defense are equally potent for offense.

The timing of Straiker’s findings coincides with a parallel investigation in South Korea, where authorities are examining similar AI-driven threats to 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. This geographic spread suggests that the vulnerability is not limited to Latin America but is a global concern for financial sectors adopting AI technologies.

In response to these growing risks, AI companies are moving to fortify their defenses. Nvidia, which has shifted its focus to hardware for AI data centers, recently launched the Open Agent Safety Platform. This reference system is designed to secure rogue AI agents through a new stack of hardware and software security. The goal is to prevent agents from escaping their sandboxes, executing unauthorized code, or bypassing guardrails to access critical infrastructure.

The rapid adoption of AI agents in cyberattacks signals a new phase in the digital arms race. As financial institutions grapple with the aftermath of these breaches, the industry faces pressure to implement more robust verification methods. The ability of a single group to automate attacks on hundreds of targets simultaneously challenges the traditional perimeter-based security models, suggesting that the next frontier of cybersecurity will be defined by how well we can contain and verify the autonomous actions of AI systems.