A gang used Claude, Grok and Kimi agents to attack more than 400 banks and payment firms, researchers say
A cybercrime group ran a team of AI agents against more than 400 financial targets across Latin America in a matter of months, according to a report published yesterday by Straiker, a US AI-security startup, and first covered by The Information. Straiker calls the group CyberCartel and says it worked as "Tr4nsHack Team" behind a front company, Hexagon Security Labs Mexico; its targets included Mexican banks, fintechs and payment providers. The system, which the gang called CARLIY0S, was built on PentAGI, an open-source tool for automated penetration testing. Straiker says a Claude agent running in the Cursor code editor acted as commander, Claude Opus models wrote exploit code, scouted targets and argued over which attack to try before an "arbiter" model picked one, and five Grok agents dug passwords and keys out of systems they had broken into. When a model refused, the work passed to Moonshot's Kimi models, and when those refused too, to attack code the gang had written in advance. The trick for getting past the models' safeguards was ordinary: the prompts described every job as an authorised security test, an academic exercise or a bug-bounty report. Straiker got a view inside after the group leaked parts of its own operation; a 24 September snapshot of its stolen-data store held 4,493 records, including 2,911 email addresses, 903 Mexican bank account numbers, 391 API keys, 20 signing keys for login tokens, six backdoors and notes on more than 200 vulnerabilities, some previously unknown. The Information reports that data was confirmed stolen from at least one institution, which is being notified; how much money, if any, was taken is not known. Anthropic says it investigated and cut off the gang's access to Claude, and xAI is investigating the Grok accounts. Straiker sells AI-security products and the findings have not been independently checked, but the method matters beyond Mexico: a few people can now run the reconnaissance and exploit-writing that used to need a team, which is why Straiker's chief executive, Ankur Shah, told The Information that "offense is way ahead of defense right now". For Kenyan banks, saccos and fintechs, the practical lessons are the old ones done faster: rotate API keys and signing keys often and give them short lives, watch for logins from unusual places even when they look local, since this gang routed traffic through residential connections, and treat an AI request framed as a "pentest" as a warning sign, not a permission slip.
Sources: Straiker · The Information · PYMNTS · DigitalToday
