CyberCartel Agents Trick AI Refusals to Steal Bank Data Across Latin America
Straiker research shows a hacking group active since 2012 bypassed model safeguards to access customer accounts.
By Carolina Wood ·
A hacking group working in Latin America has used AI agents to access bank accounts and customer data, with victims providing little human involvement in the intrusions, according to research from the cybersecurity startup Straiker. The findings were shared by The Information.
CyberCartel, the group behind the attacks, has operated since at least 2012. Straiker's research indicates that in recent months, AI tools have expanded how many organizations the group can target and the speed at which it works. Passwords and customer data have been stolen, and individuals' bank accounts have been accessed.
Agents from Grok, which is made by SpaceX, and from Claude, made by Anthropic, were among the tools used in the operation. An agent from the Chinese lab Moonshot AI was also used. Straiker gave its findings to SpaceX and Anthropic, and both companies are working to cut the group off from their services. Anthropic has blocked the group's access to Claude. Straiker is still trying to reach Moonshot AI.
The models are designed to refuse requests that would help carry out cyberattacks, but CyberCartel bypassed these checks by telling them they were operating in test environments, Straiker reported. When one model declined to continue, the group switched to another. If that model also refused, the agents were assigned to deploy malware instead.
Straiker's research does not indicate the group used Anthropic's Mythos model. Anthropic disclosed that its technology was used in a cyberattack in November 2025. Claude was later used to break into OpenAI, a rival AI developer, in an operation carried out by researchers rather than criminals.
The Straiker report follows news that South Korean authorities are investigating a similar problem at banks in that country. Hana Bank, KB Kookmin Bank, and Shinhan Bank are among those affected, and the personal credit information of at least 25,000 Shinhan Bank customers has been leaked.
Nvidia, which is now focused on hardware for AI data centers, has launched the Open Agent Safety Platform, a software platform and reference system design that uses a new hardware and software security stack to contain rogue AI agents. It is intended to help stop agents from escaping their sandboxes, executing unauthorized code, gaining unauthorized access to critical infrastructure, or bypassing guardrails.
