AI in action · Investor protection

One Model Scouted, Four Planned, One Gave Orders: How a Criminal Group Pointed AI Agents at 400 Latin American Financial Firms

A criminal group used AI agents built on Claude and Grok to target over 400 Latin American financial firms, a report said on October 21, 2026. The group, CyberCartel, went after banks, payment apps and other financial firms, The Information reported. The details come from security researchers who found the group's leaked internal files. Grok agents researched targets, four Claude agents weighed attack methods and a Claude model in a coding app directed the work; when a model refused, the task passed to a smaller one, then to ordinary malware. Passwords and customer data were stolen with little human effort, and one institution's breach was confirmed. Anthropic said it cut off the group's access.

Oct 23, 2026

The group has been attacking Latin American financial institutions since about 2012. What changed this year is who does the legwork. According to research by the security start-up Straiker, shared with The Information and reported on October 21, 2026, the group known as CyberCartel built a chain of AI agents on commercial models and used it to automate attacks on more than 400 entities in Latin America, recently including banks and payment apps. The agents reportedly stole passwords and customer data and gained access to customer bank accounts with little human involvement. The researchers could describe the system because the group accidentally leaked its own internal operational information.

The division of labour reads like an org chart. Agents running on Grok, the model now developed under SpaceX, searched the web for information about each target. Four agents on Anthropic's Claude Opus and Sonnet models then discussed which attack method suited the target best. A further Claude model, running inside the AI coding app Cursor, acted as commander and directed the overall work, according to Digital Today's account of the report. To get past safety controls, the group reportedly told the models they were working in harmless settings such as academic tests or internal corporate security checks. When Claude refused anyway, the task was handed to a smaller model, Moonshot AI's Kimi; when Kimi refused too, the system fell back on planting conventional malware.

The scale figure needs care. More than 400 is the number of entities targeted, not the number breached; the institutions were not named, and The Information said it confirmed that sensitive data had been stolen from one specific financial institution, which it did not identify because customer notifications were under way. The model developers have responded: Anthropic said it opened an investigation after receiving the findings and blocked the group's use of Claude, and SpaceX is working to block access, while the researchers were still trying to reach Moonshot. "Offense is way ahead of defense right now," Straiker's chief executive told The Information.

It follows the case of an attacker who used an off-the-shelf AI coding agent against South Korean banks, and it fits what supervisors have been saying since spring. The ECB has given eurozone banks until October 31 to file plans against AI-enabled cyberattacks; the Bank of England's deputy governor called advanced AI a systemic cyber risk earlier this month. The common thread is cost: a reconnaissance and planning job that once needed a team of skilled people now needs a few subscriptions and a coordinator, which is why the number of targets can run into the hundreds.

For customers of banks and payment apps, the breach itself is out of their hands; what follows it is not. Stolen customer data is the raw material of the next scam, usually a call or message that knows your name, your bank and the last digits of your card. A breach notification from a financial firm is a reason to change that account's password, turn on stronger login protection and treat any unexpected contact claiming to be the bank as suspect until verified through the number on the card.

Sources: PYMNTS — Hackers Employ Grok and Claude to Attack LatAm Banks (Oct 21, 2026); Digital Today — Claude commands and Grok scouts in AI hacking that hit banks, report finds (Oct 22, 2026); MarketScreener / MT Newswires — Hackers Use Claude and Grok to Automate Attacks on Latin American Banks (Oct 21, 2026)