Straiker: hackers used Claude and Grok to target Latin American banks

A long-running criminal group used several commercial AI models to go after banks and payment firms in Latin America, according to security firm Straiker. Straiker says it found CARLIY0S, a system the CyberCartel group used to run attacks on banks and payment providers across the region.

The scale stands out. Straiker says the group, which now goes by Tr4nsHack Team, went after over 400 financial targets within a few months. Straiker documents the group's tools and stolen data, but it has not shown how many of those targets were actually breached rather than probed.

Key facts

• Straiker says CARLIY0S runs on the open-source PentAGI offensive framework and uses Kimi AI's model, Claude Sonnet and Opus, and xAI's Grok.
• Straiker says a snapshot of the group's stolen-credential database held 4,493 records gathered by five Grok miners.
• Straiker says the stolen data was led by 2,911 email addresses and 903 account numbers.
• Straiker says Anthropic revoked the compromised credential used to reach Claude, and xAI is looking into the Grok-linked activity.
• Straiker says CyberCartel has been tracked since 2012.

How the system worked

CARLIY0S split the work among several AI roles. Straiker says four roles based on Claude were set up to argue over targets and hand a final call to a designated arbiter.

The group found ways around safety rules. Straiker says the attackers presented their requests as approved red-team work or capture-the-flag exercises to push models past their safety policies.

There was also a backup plan. Straiker says when fallback models turned down attack requests, the orchestrator switched to a local file of ready-made attack code.

The stolen material went beyond email. Straiker says it reviewed a snapshot of the credential database on September 24, 2026, holding 4,493 records. Straiker says the haul also included 391 general API keys and 67 passwords stored in plain text.

The access had serious potential. Straiker says the access could enable abuse of SPEI, Mexico's interbank payment system.

Background

CyberCartel is not new. Straiker says the group has been tracked since 2012 and has aimed at Latin American financial firms for more than a decade.

AI developers have been tracking similar abuse. Anthropic's September 2026 report says it disrupted misuse between December 2025 and August 2026 across seven harm areas, including cyber operations. The same report says none of those cases involved its Claude Fable or Mythos-class models, apart from one illicit distillation case.

Anthropic also sees the pattern spreading. Anthropic says an autonomous-attack model it documented in November 2025 has now spread across every class of actor it investigated. Its report put the shift plainly: "Sophisticated attacks no longer require sophisticated attackers."

Who is affected

The victims have not been named. Straiker says it anonymized the victims as Latin American and Mexican financial institutions, banking entities and fintech payment gateways.

Customers of those firms have reason to pay attention. Straiker says the stolen database included 903 account numbers, though it has not said which banks or fintechs those accounts belong to.

What it means for you

For security teams, the message is about speed. Straiker says defenders should put money into AI-driven testing, detection and response so they can match attacks run by machines.

For the AI companies, the case shows how their models get pulled into crime. Straiker says the group switched among Claude, Grok and Kimi models and fell back to stored attack code when models refused.

All of these figures come from Straiker, a security vendor, and have not been independently verified.

Practical steps

• Security teams at financial firms can review Straiker's published report on CARLIY0S for the tools and methods it describes.
• Straiker recommends investing in AI-driven testing, detection and response.
• Teams can read Anthropic's September 2026 threat intelligence report for its account of AI misuse in cyber operations.

What happens next

One AI company is still looking into its side. Straiker says xAI is still investigating the Grok-linked activity after being told before publication. xAI had not published a response in the sources reviewed.

Several facts remain unknown. It is not known which banks or fintechs were affected, or how many of the targets were breached.

FAQ

Which AI models were involved? Straiker names Claude Sonnet and Opus, xAI's Grok and Kimi AI's model.

Has access been cut off? Straiker says Anthropic revoked the compromised credential used to reach Claude, and xAI is looking into its side.

Were customer accounts exposed? Straiker says the stolen database held 4,493 records, including 903 account numbers.

How did the group get past safety rules? Straiker says it framed requests as authorized red-team work or capture-the-flag exercises.

Who is CyberCartel? Straiker says the group has been tracked since 2012 and now operates as Tr4nsHack Team.

What could the access be used for? Straiker says the access could enable abuse of SPEI, Mexico's interbank payment system.

What other data was in the database? Straiker says it held 2,911 email addresses, 391 general API keys and 67 cleartext passwords.

What happened when models refused? Straiker says the orchestrator fell back on a local file of pre-written attack code.

What framework did the group build on? Straiker says CARLIY0S is based on PentAGI, an open-source offensive framework.

Has Anthropic seen similar misuse? Anthropic says it disrupted misuse across seven harm areas, including cyber operations, between December 2025 and August 2026.