CyberCartel, a hacking group active since at least 2012, has been using commercial AI agents to attack Latin American banks and mobile payment apps with little hands-on work from its operators.
The findings come from Straiker, a cybersecurity startup. It said the group has stolen passwords and customer data and gained access to individuals' bank accounts. The agents involved include Grok, Anthropic's Claude and an agent from China's Moonshot AI. Straiker shared its research with SpaceX and Anthropic, and both are working to lock the group out. Anthropic has already blocked its access to Claude. Straiker is still trying to reach Moonshot AI.
The method is blunt. CyberCartel told the models they were running in test environments. When one model refused, the group moved to another. When that one refused too, the agents were tasked with deploying malware instead.
That is the part worth dwelling on. The safeguards did not fall to a clever exploit. They fell to a lie, and the fallback chain means a block from one provider only pushes the group to the next. Speed is the other change: Straiker's research shows the group can now hit more businesses than human-run hacking allowed.
The group does not appear to be using Mythos, the Anthropic model whose announcement came with warnings that hackers could turn such systems to exactly this kind of attack. The damage here comes from models already on the market.
It is not an isolated case. Anthropic disclosed a cyberattack carried out with its AI in November 2025, and South Korean authorities are investigating a similar problem at local banks, including Hana Bank, KB Kookmin Bank and Shinhan Bank. At least 25,000 Shinhan customers had personal credit information leaked.
Nvidia's newly launched Open Agent Safety Platform aims to stop rogue agents from escaping sandboxes or bypassing guardrails. That targets agents that misbehave on their own. It is less clear what it does when the operator is the attacker and the model has been talked into cooperating.