Straiker finds a crime group that attacks banks with AI models

Straiker, a security firm, reports that the CyberCartel group made an attack system, CARLIY0S, to hit banks and payment providers in Latin America. On 24 September 2026, Straiker agents got a copy of the database of stolen data of the group. It had 4,493 records. Straiker says this is the first such study from the own work area of a criminal group.

How it works

CARLIY0S uses PentAGI, an open-source tool for automated security tests. It gives different tasks to commercial AI models. Four Claude-based roles judge attack alternatives, argue about targets, and give a last decision through one arbiter.

The proof

The database had 2,911 email addresses and 903 account numbers. It also had 6 backdoors and more than 200 flaws, some not known before. The group told the models that the work was an approved security test.

What is not known

Straiker names the victims only as Latin American banks and payment firms. The source gives no count of attacks that worked. The report says the group presumably has ties to Mexican cartels.

Insight: Criminals leave their AI work open for defenders to read

Straiker got the database of the group. In a Korean bank attack, CrowdStrike used open Claude Code session records to connect the breaches. NorForge found more than 20 open panels of the Cairn tool. Attackers use the AI services of model makers. The makers hold the stop control, and the open records seem to give defenders a rare view. Maybe the model maker is the best position to stop a group.