Hackers used Claude, Grok to attack banks, Straiker says

A long-running Latin American cybercrime group used a network of AI agents to break into banks and payment apps with very little human direction, The Information reported on October 21. The AI agents ran on models from Anthropic, SpaceX and the Chinese lab Moonshot AI.marketscreener.com

According to The Information, the group, known as CyberCartel, aimed its automated attacks at more than 400 financial targets in Latin America. The findings come from an investigation by the cybersecurity startup Straiker, which shared its research exclusively with the outlet. The Information called it one of the first detailed reports of criminals using AI models to successfully automate break-ins at financial institutions.theinformation.com

What the agents did

The targets included several banks and payment apps. The AI agents could steal passwords and customer data and get into users' bank accounts on their own, with almost no human involvement, according to a summary of the report published by the Chinese outlet Jinse Finance. MT Newswires also picked up The Information's report. The headline names Grok, the chatbot developed by xAI, which is now part of SpaceX.marketscreener.com+1

Moonshot is already under scrutiny in Washington. In September, the NSA, FBI and CISA accused it and five other Chinese AI companies of pulling billions of tokens from American models, including Claude and Grok, to train their own systems, CoinDesk reported.coindesk.com

A familiar group with new tools

CyberCartel is not a new player. IBM security researchers say the group has been active in Latin America since 2012. It has mostly rented existing malware rather than building its own, using it against government offices and financial institutions. Its methods have included fake billing and government websites and malicious browser extensions that steal banking logins. Darktrace has linked the group to campaigns aimed at bank and tax-service users, especially in Mexico and Chile. It used malvertising and phishing pages that imitated real institutions.ibm.com+1

Rising concern over agent attacks

Straiker sells security tools for AI agents. In July it published a threat report describing what it calls "AI-powered persistent threats": attackers that are themselves agents and can automate reconnaissance, write exploits and keep long-term access to systems, according to the company's STAR Labs report. The company recently raised $64 million in a Series A round, The Next Web reported.straiker.ai+1

The new findings point to a problem for defenders. Tools that attackers could once run only a few at a time can now be copied across hundreds of targets at once. In separate research, Darktrace said AI-assisted attacks "still leave a behavioral trace." It added that while AI can shape how attacks are built and scaled, "it does not make them operationally invisible."darktrace.com