CyberCartel Deploys AI Agents to Accelerate Attacks on Latin American Finance
A long-standing hacking collective has integrated artificial intelligence tools to automate breaches of banks and payment apps across the region.
By Tunde Adeyemi ·
The hacking collective known as CyberCartel has begun using artificial intelligence agents to automate its intrusions into Latin American financial institutions and mobile payment applications. This development marks a significant shift in tactics for a group that has operated in the region for over a decade. According to recent research by the cybersecurity startup Straiker, the integration of AI has allowed the group to expand its reach and increase the speed of its operations far beyond what human-driven efforts could achieve.
The adoption of these automated tools has resulted in substantial data compromises. Straiker notes that the group has successfully stolen passwords and customer records while accessing individual bank accounts with minimal direct human intervention. The increased efficiency of the AI agents enables the attackers to target a larger number of organizations simultaneously, putting significant pressure on security teams in the region to defend against a faster and more distributed threat.
The operation relies on a mix of AI models from major global developers. CyberCartel has utilized agents from SpaceX, specifically the tool known as Grok, and agents from Anthropic, including Claude. Additionally, an agent from the Chinese laboratory Moonshot AI has been part of the attack framework. Straiker has shared its findings with both SpaceX and Anthropic. While Anthropic has already blocked the group's access to Claude, the companies are working to sever the attackers' connections to their respective services. Straiker continues to seek engagement from Moonshot AI regarding the incident.
The case highlights vulnerabilities in the safety mechanisms built into modern AI models. These systems are generally designed to refuse requests that assist in cyberattacks, but the hackers reportedly circumvented these safeguards by claiming the models were operating in test environments. When one model declined to proceed, the attackers switched to a different provider. If that model also refused, the agents were reassigned to deploy malware instead. This adaptive approach demonstrates how attackers can leverage redundancy in AI services to maintain operational momentum.
It is worth noting that Straiker's research does not indicate that the group used Anthropic's Mythos model. This distinction is relevant because Anthropic drew attention earlier in the year when it unveiled Mythos and warned that hackers could exploit it and similar models for this specific type of attack. The release of Mythos had previously prompted security teams worldwide to address thousands of vulnerabilities that the model had identified, underscoring the dual-use nature of advanced AI capabilities.
This is not the first time Anthropic has discovered that its technology is being used by malicious actors. In November 2025, the company disclosed details of its AI being employed in a cyberattack. Subsequently, the Claude model was used to breach the security of OpenAI, a rival developer. However, that specific operation was conducted by researchers rather than criminal groups, suggesting a broader pattern of AI utilization in both offensive security and criminal hacking.
The timing of the Straiker report coincides with emerging concerns in Asia. South Korean authorities are currently investigating a similar issue affecting local banks. Major institutions including Hana Bank, KB Kookmin Bank, and Shinhan Bank have been impacted. Personal credit information belonging to at least 25,000 Shinhan Bank customers has been leaked in this separate but related wave of incidents. This parallel development suggests that the use of AI in financial hacking is a growing global trend rather than a regional anomaly.
Industry leaders are increasingly aware of the risks posed by autonomous agents. Nvidia, which is currently focused on hardware for AI data centers, recently launched the Open Agent Safety Platform. This software platform and reference system design utilizes a new hardware and software security stack to contain rogue agents. The system is intended to prevent agents from escaping their sandboxes, executing unauthorized code, or bypassing established guardrails. These efforts reflect a broader push to secure the infrastructure that supports the next generation of digital interactions.
Source: Tom's Hardware
