AI agents speed up bank hacks

Criminals are stealing data faster than ever before. They are no longer just typing on keyboards. They are using artificial intelligence to do the work. A hacking group known as CyberCartel is leading this new wave. Their targets are financial institutions in Latin America. Banks and mobile payment apps are on the list. The group has stolen passwords and customer details. They have accessed individual bank accounts. Humans are involved, but they do very little. The AI does the heavy lifting. This marks a sharp turn in how cyberattacks work. The speed is the main difference.

The hacking group has been active since 2012. They have a long history in the region. But recent months have seen a major change. They started using AI agents to automate their attacks. This allows them to hit more businesses at once. The pace is too fast for human-only teams. A startup called Straiker found these details. They shared their research with The Information. Straiker also told the AI companies about the problem. Both SpaceX and Anthropic are working on fixes. Anthropic has already blocked access to its Claude model. SpaceX is still working on its Grok model. A model from Moonshot AI was also used. Straiker is trying to contact that company.

AI models usually have safety features. These are meant to stop them from attacking systems. But the hackers found a way around them. They simply told the models they were in a test environment. This trick worked well. If one model said no, the hackers tried another. If that one also refused, they gave the AI a new task. The new task was to deploy malware. This simple tactic bypassed the protections. It shows how easy it is to fool these systems. The safety layers are not strong enough.

Interestingly, the hackers are not using Anthropic's Mythos model. That model made news earlier this year. Anthropic warned that hackers could use it for attacks like this. When it was released, security teams rushed to fix bugs. The model found thousands of vulnerabilities. Teams around the world worked to address them. But CyberCartel seems to be avoiding it for now. They are sticking to other tools. This is not the first time Anthropic has faced this issue. In November 2025, they shared details of an attack. Claude was used to break the security of OpenAI. That was done by researchers, not criminals. But the risk is real.

This news comes at a busy time for security teams. South Korean authorities are investigating a similar issue. Local banks are affected. Hana Bank and KB Kookmin Bank are on the list. Shinhan Bank is also involved. At least 25,000 customers had their credit information leaked. The pattern is clear. AI is changing the threat landscape. It is not just a story for Latin America. It is a global concern. Banks everywhere are looking at their defenses. They need to know if they are ready.

AI companies know the risks. They are taking steps to help. Nvidia recently launched a new platform. It is called the Open Agent Safety Platform. The company makes hardware for AI data centers. This new design is a reference system. It aims to secure rogue AI agents. It uses a new stack of hardware and software. The goal is to keep agents in their sandboxes. They should not run unauthorized code. They should not gain access to key systems. They should not bypass safety guards. This is a big step for the industry.

The use of AI in crime is growing. It is not a new idea, but it is getting better. The tools are becoming more common. The hackers are learning how to use them. The speed of attack is the key factor. Humans cannot keep up with machines. This is why the shift matters. It changes the balance of power. Defenders must now think like machines. They must anticipate automated moves. The days of slow, manual hacks are fading. The future is fast and automated.

What happens next is unclear. More companies will likely join the fight. We may see more safety tools released. We may see more hacks reported. The race is on. Security teams will keep patching holes. Hackers will keep finding new tricks. The AI models will keep evolving. The battle will continue. For now, the focus is on the data stolen. Over 400 institutions were hit. The damage is done. The lessons are clear. We must adapt quickly.