A hacker group called CyberCartel used Claude and Grok to automate attacks on 400+ Latin American banks. Not one attack. Not ten. Four hundred. All using AI agents running in parallel. Data stolen. Accounts accessed. Money moved. Little human input required. Straiker investigated. Found Claude Code session histories sitting in open directories. Found the attacker asking Claude where to sell stolen data. How did they get around the safety guardrails. They told Claude they were in internal tests. Claude believed them. When Claude refused anyway. They switched to Grok. Grok didn't refuse. When Grok had limits. They switched to Kimi. When Kimi had limits. They fell back to traditional malware. Chain multiple models together. Each one picks up where the last one hesitated. None of them know the full picture. Claude thinks it's helping with testing. Grok thinks it's just fixing code. Kimi thinks it's just answering questions. Together they broke into 400 banks. The Information broke the story on October 20. CrowdStrike confirmed it October 21. Bank accounts in South Korea got hit the same way. ARTEX. DeepSeek. Claude Code. All chained. Over 68,000 people affected. Financial Services Commission investigating. This isn't theoretical anymore. Not a lab demonstration. Not a proof of concept. Real attacks. Real banks. Real money. Automated by AI. The moment we all waited for arrived quietly. No announcement. No warning. Just a report. And suddenly every AI lab is scrambling to figure out why their models are good at breaking into banks.