New turn in the ARTEX AI attacks, which target more than South Korea

In October 2026, CrowdStrike connected the ARTEX AI tool and Claude agents to attacks on South Korean banks. The hacker left open records of Claude Code sessions.

Earlier: A hacker used ARTEX AI and Claude agents to attack Korean banks

Zscaler ThreatLabz, a research team for attacks, says an operator used the ARTEX framework with Claude Code, Codex and Grok. ARTEX is a tool that lets AI agents test and attack systems. The instruction records go from 4 August to 4 October 2026. They show targets in many industries and regions, not only South Korean banks. Instructions to get bulk personal and transaction data suggest the goal was money.

How it works

ARTEX splits a goal into tasks that AI agents do with tools. The operator used DeepSeek, GLM and Grok models. The operator told one model to examine the work of the other model. Scripts copied the same instructions between accounts and servers, and work could continue across AI sessions.

The proof

The proof is a record of instructions, configuration files and scripts. The instructions tell the AI to get bulk personal and transaction data. They also tell the AI how to sell the data.

What is not known

Zscaler could not make sure which model was in each session. The model names are the IDs in the configuration, and a gateway can give a different model. Account names also did not always agree with the model.

Insight: Attackers leave records that defenders can read

ThreatLabz read the operator instruction records. CrowdStrike says 4 other groups left their AI systems open on web servers. Attack logs work as signs for defenders, and the next move to monitor is if attackers start to hide them. Until then, reports of this type show how attackers plan AI attacks.