CyberCartel: AI Automation Reshapes Digital Bank Heists in Latin America
A long-standing hacking group is leveraging autonomous AI tools to accelerate data theft and account access across regional financial institutions.
By Luana Sato ·
The hacking collective known as CyberCartel has fundamentally altered its operational playbook by deploying artificial intelligence agents to automate attacks on Latin American banks and mobile payment applications. This strategic shift, identified by the cybersecurity startup Straiker and reported by The Information, marks a significant escalation for a group that has been active in the region since at least 2012. The adoption of autonomous tools allows the criminal organization to expand its reach and speed far beyond the capabilities of human-driven operations, fundamentally changing the landscape of digital financial crime in the area.
The quantitative impact of this new approach is substantial. Straiker’s research indicates that in recent months, the use of AI has allowed CyberCartel to target a wider array of organizations at a pace that manual hacking cannot sustain. This efficiency has led to the theft of passwords and customer data, as well as unauthorized access to individual bank accounts. Notably, these intrusions require minimal direct involvement from the criminals themselves, suggesting a move toward a largely hands-off execution model where the software handles the bulk of the intrusion process.
The technical infrastructure behind this operation involves a mix of major AI providers. CyberCartel has utilized agents from SpaceX, specifically the Grok model, alongside Anthropic’s Claude. The group has also integrated an agent from the Chinese laboratory Moonshot AI. In response to these findings, Straiker shared its data with both SpaceX and Anthropic. The two U.S. companies are actively working to sever the group’s access to their services, with Anthropic already blocking its access to Claude. Straiker is currently in the process of contacting Moonshot AI regarding the breach.
The incident highlights the current fragility of safety mechanisms built into large language models. These systems are generally designed to reject requests that facilitate cyberattacks, yet CyberCartel reportedly bypassed these safeguards by claiming they were operating in test environments. When one model refused to continue, the hackers would switch to another. If that model also declined, the agents were reassigned to deploy malware. This adaptive strategy demonstrates how attackers can exploit logical gaps in AI reasoning to maintain their operations.
It is important to note that Straiker’s report does not suggest the group used Anthropic’s newer Mythos model. Anthropic had previously drawn attention to this model, warning that hackers could exploit it for similar attacks. The release of Mythos had already prompted security teams globally to address thousands of vulnerabilities that the model had identified. This current event is not the first time Anthropic has dealt with bad actors using its technology. In November 2025, the company disclosed details of its AI being used in a cyberattack. Later, Claude was used by researchers, not criminals, to break the security of OpenAI, a rival AI developer.
The timing of this report is significant as it follows closely on the heels of reports from South Korea. Authorities there are investigating a comparable issue affecting major banks, including Hana Bank, KB Kookmin Bank, and Shinhan Bank. The personal credit information of at least 25,000 Shinhan Bank customers has been leaked in that incident. These parallel developments suggest a growing trend in the use of AI for financial sector attacks across different continents.
Industry leaders are already responding to these emerging threats. Nvidia, which is currently focused on hardware for AI data centers, recently launched the Open Agent Safety Platform. This new software platform and reference system design utilizes a combined hardware and software security stack to contain rogue AI agents. The platform is intended to prevent agents from escaping their sandboxes, executing unauthorized code, gaining access to critical infrastructure, or bypassing standard guardrails, aiming to create a more robust defense against automated digital threats.
