New turn in ARTEX attacks as Hunt.io shows the hacker servers

In October 2026, CrowdStrike connected the attack on South Korean banks to open directories with records of the hacker work in Claude Code. ARTEX is an AI tool that tests and attacks systems.

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

Hunt.io, a company that finds attacks, made a study of the servers behind the ARTEX attacks. Seven South Korean financial institutions had data exposed from September to October 2026. About 66,000 to 67,000 persons and 2,200 corporations were in the data. The hacker started Claude Code software 510 times since 4 August, but it sent the work to DeepSeek, not to Anthropic.

How it works

Small proxy scripts made Claude Code think that it spoke to an Anthropic server. The scripts gave an incorrect set of models and sent each message to DeepSeek. The path went through Claude Code Router, an open-source project.

The same record shows 19 custom skills, which are sets of instructions that the agent can use again.

The numbers

Hunt.io looked for ARTEX signs on the internet:

• An ARTEX certificate issuer gave 19 addresses.
• A malware class gave 1,303 addresses in 30 countries.
• These show where addresses are registered, not where the hacker sits.

What is not known

Hunt.io cannot make sure that other tools in the files, for example tools for a cryptocurrency market, connect to the attacks on the banks. CrowdStrike thinks, with moderate confidence, that one Chinese-speaking person who wanted money did the attack.

Insight: Removing a tool from GitHub does not stop its users

Genians says the ARTEX code and Docker image were removed on 11 October, but copies stay. Check Point Research found 1,608 live ARTEX control panels in November. The hacker used a usual coding agent and changed the model behind it, and the model maker had no record of it. It seems the control point is the open server, not the model maker. Defenders must look at these servers.