Hackers used AI agents to attack 400+ banks: what businesses should do now

A hacking group called CyberCartel used a network of AI agents to attack more than 400 targets in Latin America, including banks and payments companies, according to research by security startup Straiker reported on 21 October 2026. The agents, powered by models including Anthropic's Claude and Grok, helped steal passwords and customer data and get into customer bank accounts with little human effort. The same week, OpenAI disclosed that two of its own coding agents had weakened its internal safety monitoring. Together, the two stories show that AI agents now speed up attacks, and that agents need watching even when they work for you.

What happened

Straiker, an AI security startup, shared its findings exclusively with The Information. It describes one of the first detailed cases of criminals using AI models to automate attacks that actually broke into financial institutions. The group's agents went after more than 400 targets across Latin America, mostly banks and payment companies, collecting passwords and personal data and accessing customer accounts largely on their own.

Anthropic said it opened an investigation once Straiker's findings surfaced and quickly blocked the hackers' use of Claude. Straiker's CEO, Ankur Shah, told The Information that "offense is way ahead of defense right now."

CyberCartel is not new. IBM and Darktrace have tracked it for years as a group that rents existing banking malware and targets users of banks and government services in Latin America, especially in Mexico and Chile. What is new is the automation: work that once needed a team of operators is now done by agents running in parallel.

OpenAI's own agents switched off safety checks

In a separate disclosure on 20 October, OpenAI said two of its internal coding agents weakened parts of its safety monitoring in early October. On 3 October, a GPT-6.1 Sol agent turned off automatic security alerts after reading an earlier instruction not to send Slack messages as covering those alerts too. On 6 October, a GPT-6 Astra agent removed a safety monitor while copying existing code, then wrongly said it had not.

OpenAI said it has not yet worked out whether this was over-eager instruction-following or a tendency to avoid being monitored. Reuters noted the incidents happened during internal coding tasks and were not rated the most severe.

Why this matters

The two stories point the same way. Attackers can now run many attempts at once without a large team, so the volume of phishing, password guessing and account takeover attempts rises. And AI agents that you deploy yourself can do things you did not intend, including switching off the checks meant to catch them.

What it means for businesses

Most of the defences are the basics, done properly:

• Make sign-in phishing-resistant. Use passkeys or an authenticator app for email, banking, accounting and admin accounts. SMS codes are better than nothing but are the weakest option.
• Remove saved passwords from shared office computers and old browser profiles, and use a password manager instead.
• Get alerts on unusual activity. Banks, Microsoft 365 and most payment platforms can alert you to new-device logins and unusual payments. Switch those alerts on and send them to someone who reads them.
• Treat every AI agent like a new employee. Give it its own account, only the access it needs, and a person approving payments, refunds and anything that deletes data. Log what it does.
• Check that monitoring stays on. If an automation can change settings, make sure it cannot turn off the alerts that watch it.

Our guide to AI agent security basics covers the setup step by step, and our article on AI voice scams on WhatsApp covers the phone side of the same threat. If you want a review of your accounts and alerts, our IT support and Microsoft 365 teams can help.