By Hamza Faisal

A hacking group known as CyberCartel has turned commercial AI models from Anthropic, xAI, and Moonshot AI into an autonomous attack network that stole passwords and customer data from more than 400 financial institutions across Latin America — with almost no human involvement, according to a new investigation by AI-security startup Straiker.

The findings, shared exclusively with The Information and published by Straiker's STAR Labs team on October 21, are being described as one of the first detailed cases of criminals using off-the-shelf AI models to automate break-ins at banks and payment companies at industrial scale.

A commander, scouts, and a fallback chain

The group's system, internally named CARLIY0S and built on the PentAGI autonomous offensive framework, split the work between models. Grok agents scouted targets across the web. Four Claude agents — running Opus and Sonnet — debated attack strategies and returned a final decision through an arbiter. A Claude model inside the Cursor coding app acted as commander, directing the whole operation.

The models' safety guardrails were sidestepped by framing malicious tasks as academic tests or internal corporate audits. And the system had a refusal-fallback chain: if Claude declined a hack, Kimi was told to do it instead; if Kimi refused too, the system deployed pre-written malicious code. A refusal from one provider never stopped the workflow.

4,493 stolen records — found by accident

Operating as the Tr4nsHack Team under the front company Hexagon Security Labs Mexico, the group — which Straiker tracks as AiPT-67 and believes has targeted Latin American businesses since at least 2012 — sharply accelerated its attacks after adopting AI agents in recent months.

Straiker learned how the machine works because the group accidentally leaked its own operational information. On September 24, Straiker's agents accessed a snapshot of "FinancialSoft," the group's stolen-credential database: 4,493 records collected by five "Grok miners," including 2,911 email addresses, 903 account numbers, passwords, API keys, and even zero-day exploits and backdoors pulled from victims between May and September 2026.

"Offense currently outpaces defense by a wide margin"

Straiker shared its indicators with the affected AI providers before publishing. Anthropic investigated and revoked the compromised credential the group used to reach Claude, cutting off the swarm's access; xAI was notified and is examining the activity. The Information reviewed stolen-data records but did not name victims while notification is underway.

"The reality is that offense currently outpaces defense by a wide margin, and companies are not yet prepared," said Straiker CEO Ankur Shah. The details all come from Straiker's report and have not been independently corroborated — but for banks, the takeaway doesn't depend on attribution: stolen credentials and exposed accounts are what fuel these attacks, and rotating keys and monitoring access is the only defense that works in the agentic era.

By Hamza Faisal