Hackers used Claude and Grok agents against 400+ Latin American banks, report says
Security firm Straiker says a group called CyberCartel ran AI agents on Claude and Grok to break into banks and payment apps. Here is what the report found.
On October 21, 2026, security firm Straiker reported that a criminal group called CyberCartel ran a network of AI agents built on Anthropic’s Claude and Grok to attack more than 400 banks, payment companies and other financial institutions in Latin America. Claude agents planned and wrote the attacks; Grok agents collected stolen passwords and keys. Anthropic says it blocked the access; SpaceXSI (formerly xAI) is investigating, according to Straiker.
What Straiker found
Straiker calls the system CARLIY0S. It was built on an open-source framework for automated security testing and used to break into banks and payment providers, with a focus that included Mexico’s SPEI payment system. Victim data in the group’s database runs from May to September 2026. The research was first shared with The Information and then published on Straiker’s blog.
How Claude and Grok were split
- Claude: several Claude-based agents did the planning. One acted as the commander, others wrote exploit code, scouted targets and argued over options before a final agent picked the plan.
- Grok: five agents Straiker calls “Grok miners” gathered secrets, such as passwords and API keys, into a database of stolen credentials.
- Kimi: models from Chinese company Moonshot were set up as a backup. When they refused offensive requests, the system fell back to attack code the group had written in advance.
What was taken
A snapshot of the group’s database from September 24, 2026 held 4,493 records, according to Straiker, including:
- 2,911 email addresses and 903 account numbers
- 391 API keys and 67 passwords stored in readable text
- 20 keys used to create login tokens, plus more than 200 known security holes, unpublished exploits and backdoors
Press reports also describe access to customer bank accounts. Straiker does not name the institutions.
What Anthropic and SpaceXSI said
Straiker says it shared its findings with both companies before publishing. Anthropic revoked the account used to reach Claude. SpaceXSI said it is looking into activity linked to the five Grok agents. Neither company has published its own statement on the case as of October 25.
What it means if you use Grok
This is a case of criminals using AI agents, not a breach of Grok itself, and nothing suggests Grok users’ own data was exposed. It does show why agent tools now come with approval cards, blocked sites and usage limits: the same features that let a Grok Bot sign in and act for you can be turned toward abuse. If you connect Grok Bot to work accounts, give each Bot only the access it needs and keep approvals on for anything that sends money or data. More background in our Grok controversies timeline and Grok privacy guide.