Hackers Use AI Agents To Steal Bank Data

A hacking group known for targeting Latin American businesses has turned to artificial intelligence tools to automate attacks on banks and mobile payment applications, according to research published Thursday by security startup Straiker. The findings were shared with The Information.

The group, which has operated as a threat since at least 2012, has recently upgraded its tactics by employing AI agents, including Anthropic's Claude and SpaceX's Grok. This shift has allowed the organization to attack a higher volume of businesses at a speed previously unattainable with human-powered hacks. Passwords and customer data were stolen, and bank accounts were accessed with minimal effort from the criminals.

The research indicates that CyberCartel has targeted over 400 institutions, including banks and payment apps. Straiker has shared its findings with SpaceX and Anthropic, and both companies are working to block the group's access to their services. Anthropic has already successfully restricted access to Claude. The hacking group has also utilized an agent from Chinese lab Moonshot AI, though Straiker is still attempting to contact that company.

While AI models typically include protections against cyberattacks, these safeguards proved easy for the group to bypass. The hackers reportedly claimed to be operating within test environments to circumvent restrictions. If a model refused to proceed, the group would switch to a different model. If that model also refused, they would task the agents with deploying malware instead.

This incident follows other reports of AI misuse. In November 2025, Anthropic disclosed that its AI was used to execute a cyberattack. Additionally, South Korean authorities are currently investigating a similar issue affecting local banks, where at least 25,000 Shinhan Bank customers had personal credit information leaked. Nvidia has also launched the Open Agent Safety Platform to secure rogue agents and prevent them from bypassing guardrails or executing unauthorized code.

Anthropic's Mythos model, released earlier this year with warnings about potential misuse, does not appear to be involved in these recent attacks.