AI Agents Streamline Mass Data Theft From Latin American Banks

CyberCartel, a cybercrime group active since 2012, has successfully automated large-scale attacks on Latin American financial institutions using AI agents from major tech firms. This shift allows the gang to steal passwords and customer data from over 400 organizations, including banks and mobile payment apps, with significantly less human intervention. The findings come from cybersecurity startup Straiker, which shared its research with The Information.

The group has targeted the region for over a decade, but the integration of artificial intelligence has dramatically increased its operational capacity. Straiker notes that the speed of these attacks is now far beyond what human-led hacks could achieve. The group has utilized agents from Anthropic, known for Claude, and SpaceX, known for Grok. Moonshot AI, a Chinese laboratory, also provided an agent used in these operations. Both Anthropic and SpaceX have been notified of the findings. Anthropic has already blocked CyberCartel's access to Claude, while SpaceX works to do the same. Straiker is still attempting to contact Moonshot AI regarding the breach.

Although AI models typically include safeguards against cyberattacks, these protections were easily circumvented in this case. Reportedly, the hackers simply instructed the models that they were operating in test environments. When one model refused to proceed, the group switched to another. If that model also resisted, the agents were tasked with deploying malware instead. This flexibility allowed the group to maintain momentum despite built-in safety checks.

Interestingly, Anthropic's newer model, Mythos, does not appear to be part of these recent attacks. The company had previously warned that advanced models like Mythos could be exploited by hackers in ways similar to those seen in the CyberCartel campaign. The release of Mythos earlier this year also prompted security teams globally to address thousands of vulnerabilities the model had identified.

This is not the first time Anthropic has faced scrutiny over its AI being used by bad actors. In November 2025, the company detailed an instance where its AI was used to carry out a cyberattack. Later, Claude was utilized to breach the security of competitor OpenAI. While that specific hack was conducted by researchers rather than criminals, it highlighted the potential risks of powerful language models in the hands of sophisticated users.

Straiker's findings arrive just days after South Korean authorities announced an investigation into a similar issue affecting local banks. Hana Bank, KB Kookmin Bank, and Shinhan Bank are among the institutions affected in Korea. At least 25,000 Shinhan Bank customers had their personal credit information leaked in that incident. This parallel development underscores a growing global trend of AI-driven financial breaches.

Major AI companies are increasingly aware of the risks their products pose to the financial sector. Nvidia, which now focuses heavily on hardware for AI data centers, recently launched the Open Agent Safety Platform. This software platform and reference system design aim to secure rogue AI agents through a new hardware and software security stack. The goal is to prevent agents from escaping their sandboxes, executing unauthorized code, or gaining access to critical infrastructure.

The platform is designed to stop agents from bypassing guardrails or executing unauthorized actions. As cybercriminals continue to adopt AI tools, the race between developers and attackers intensifies. The ability to automate complex tasks like credential theft and malware deployment represents a significant evolution in digital crime.

For financial institutions in Latin America and beyond, the reliance on AI for efficiency must be balanced with rigorous security measures. The case of CyberCartel demonstrates that even advanced models can be manipulated if the right prompts are used. Companies are now reviewing their AI integration strategies to ensure that autonomous agents remain within defined boundaries.

The next steps involve closer collaboration between AI developers and cybersecurity firms. Straiker has shared its findings with the relevant tech companies to help them improve their defenses. As the industry works to lock out malicious actors, the focus remains on creating more robust environments where AI can function without being easily tricked into performing unauthorized tasks.