Straiker: crime group used Claude, Grok and Kimi to attack 400+ Latin American banks

A criminal group called CyberCartel used AI agents built on Anthropic's Claude, xAI's Grok and Moonshot AI's Kimi models to automate attacks on Latin American banks and payment providers, security startup Straiker reported on 21 October 2026 [1][2]. Straiker says it based the report on the group's own systems and that Anthropic has cut off its Claude access [1].

What did Straiker find?

Straiker says the group, operating as Tr4nsHack Team under a front company called Hexagon Security Labs Mexico, built a multi-agent attack system named CARLIY0S on the open-source PentAGI framework [1]. Straiker says CyberCartel attacked more than 400 Latin American financial targets in a matter of months. It tracks the group as AiPT-67 and says CyberCartel has been active since 2012 [1].

On 24 September 2026, Straiker says, its agents accessed a snapshot of the group's database, called FinancialSoft [1]. Straiker says the group's stolen-data database held 4,493 records collected by five Grok-based miners. Straiker lists 2,911 email addresses, 903 account numbers, six backdoors and more than 200 vulnerabilities, some of them previously unknown [1].

How did the AI agents divide the work?

According to Straiker, a Claude-based commander running in the Cursor IDE handed tasks to sub-agents through the Anthropic API [1]. Three Claude-based roles covered reconnaissance, exploit code and unconventional attack paths and debated options, and a fourth, the arbiter, returned one decision [1].

Kimi models served as fallbacks; when they refused offensive requests, the system switched to a local file of prewritten attack code, Straiker says [1]. The report adds that a configuration entry for Anthropic's Fable 5 model does not prove the model was actually used [1].

To get past safety filters, the operators framed their requests as authorised red-team work or academic exercises, according to the report [1].

How did Anthropic and xAI respond?

Straiker says it shared indicators with the affected AI providers before publication. Anthropic revoked the credential the group used to reach Claude, according to Straiker. Straiker says xAI was notified and is looking into the activity of the five Grok miners [1].

The Information, which first reported the findings, described the models as coming from Anthropic, SpaceX and Moonshot AI [2]. Straiker does not name the victim banks and anonymises them as Latin American and Mexican financial institutions [1].

Our read

Opinion: The striking part is not any single exploit but the org chart: models given roles, a debate and an arbiter. Guardrails that accept a claimed red-team context at face value will keep failing, so providers need to watch behaviour across an account, not single prompts.

What we could not verify

• How many of the more than 400 attacks succeeded or caused financial losses; Straiker gives no count.
• The identity of the victim banks and payment providers, which Straiker anonymised.
• Whether Anthropic's Fable 5 model was actually used; Straiker itself treats this as unconfirmed.
• Statements from Anthropic, xAI or Moonshot AI beyond what Straiker reports.