AI Automation: CyberCartel's New Digital Arms Race

A persistent tension has emerged in the global cybersecurity landscape: the very tools designed to streamline digital operations are now being weaponized to automate large-scale breaches. CyberCartel, a hacking collective active since 2012, has pivoted to using autonomous AI agents to target financial institutions across Latin America. This shift allows the group to execute attacks with a volume and velocity that human-led operations could never match, raising urgent questions about the resilience of current AI guardrails against sophisticated bad actors.

The timeline of this escalation is relatively recent, though the group’s history is long. While CyberCartel has been a known threat for over a decade, its recent adoption of AI agents marked a significant leap in capability. According to research from the cybersecurity firm Straiker, the group has been leveraging specific large language models to streamline the theft of passwords and customer data. This data, including access to individual bank accounts, has been extracted from more than 400 institutions, ranging from traditional banks to mobile payment applications.

The core of the operation relies on a deceptive tactic that exploits existing model protections. Straiker’s findings indicate that CyberCartel simply instructs the AI models that they are operating within test environments. When one model refuses to proceed with a task, the hackers switch to a different provider or re-task the agent to deploy malware. This cat-and-mouse dynamic suggests that while AI models have safety features, they are surprisingly porous when faced with coordinated, multi-model attacks.

The AI providers involved are currently responding to these findings. Straiker shared its research with Anthropic and SpaceX, the latter of which develops the Grok model used in the attacks. Anthropic has already succeeded in blocking CyberCartel’s access to its Claude model, while SpaceX works to do the same. The group has also utilized an agent from the Chinese lab Moonshot AI, with which Straiker is still attempting to establish contact to discuss the breach.

Notably, Anthropic’s newer Mythos model, which was released earlier this year with specific warnings about its potential for offensive use, does not appear to be part of CyberCartel’s current toolkit. This is significant because Mythos had previously forced global security teams to address thousands of vulnerabilities it identified. This incident follows a pattern where Anthropic’s tools have been exploited by outsiders; in November 2025, the company disclosed that its AI was used in a cyberattack, and later, researchers used Claude to breach the security of its competitor, OpenAI.

The situation in Latin America mirrors recent developments in Asia, where South Korean authorities are investigating similar AI-driven attacks on local banks. Institutions including Hana Bank, KB Kookmin Bank, and Shinhan Bank have been affected, with at least 25,000 Shinhan customers reportedly having their credit information leaked. These parallel events suggest a growing global trend of AI-assisted financial crime.

In response to these emerging threats, the industry is moving toward stricter containment. Nvidia, which now focuses heavily on AI hardware, recently launched the Open Agent Safety Platform. This new hardware and software stack is designed to secure rogue AI agents, aiming to prevent them from escaping sandboxes, executing unauthorized code, or bypassing established guardrails. As AI agents become more autonomous, the race to secure them is becoming just as critical as the race to build them.