CrowdStrike names ARTEX, Chinese AI pentest agent, in Korean bank hacks

An attack chain still carrying a human at every step is now a planning assumption that no longer holds in finance. CrowdStrike reported on 7 October that an unattributed threat actor used ARTEX, an open-source Chinese agentic pentesting tool, to breach several South Korean financial organizations between late September and early October, driving the attack through multiple large language models from a single operator in Hong Kong.

What ARTEX is, and what the operator ran it on

CrowdStrike describes ARTEX as "a recently released open-source agentic penetration testing (pentesting) tool developed in China." It did not name the repository. The attacker's own instance of ARTEX ran at 38.244.50[.]120 and queried four models during the campaign: DeepSeek v4.1-flash as the primary backend, GLM-5.3 from Zhipu AI, Grok 4.6, and Claude, through an API proxy that CrowdStrike labelled "likely LLM API proxy/reseller xcai[.]pro."

CrowdStrike names two kinds of system that were hit: a loan progress inquiry service at one bank and an employee mobile work-support system at another. The number of organizations affected remains unconfirmed. CrowdStrike lists a Hong Kong IP address as the primary control centre and nine further proxy IPs the agent routed through.

What the operator asked the models to do

A chat trail recovered by CrowdStrike puts the models in two separate roles. The pentesting work went to DeepSeek, GLM and Grok through ARTEX. Claude was used by the operator for tasks the agent did not do on its own: writing a security researcher résumé that listed operational results, asking where threat actors sell Korean breach data, and asking how to find Korean Telegram data sales groups. CrowdStrike also records Claude prompts for vulnerability research against Telegram-based NFT marketplaces.

The usable part is the shape of the campaign

The report attributes the actor "with moderate confidence" as "likely a Chinese speaker and financially motivated" and lists a possible identity that CrowdStrike says it cannot definitively associate with the operator. No CVE is named. No specific technical features of ARTEX are listed.

The usable piece is the shape of the campaign. One operator drove an agent through nine proxy IPs, four model backends and a vendor API proxy against two kinds of Korean financial system over about two weeks. A rule that assumes a human types every search and reads every result still blocks the exfiltration if the exfiltration is noisy, and misses the recon if the recon arrives as patient, ordinary-looking traffic from many IP addresses. A rule that watches for an agent calling a known LLM API proxy from a surface the vendor pays for is the lever CrowdStrike's indicators give you.

Source

CrowdStrike's report is at Unknown Threat Actor Uses AI-Driven ARTEX to Target South Korean Finance. A paywalled summary of the report ran in The Information.