The Pulse
CrowdStrike Finds ARTEX in South Korean Bank Intrusions
CrowdStrike says an unidentified actor used the ARTEX penetration-testing agent and large language models in intrusions targeting South Korean financial organizations. The company assesses with moderate confidence that the actor is likely C
AI.info Team · Written with AI ·
CrowdStrike has evidence that an attacker used AI tools while targeting South Korean financial organizations, but it stops short of identifying who was behind the activity. In a report published October 7, the cybersecurity company said it found ARTEX configuration files and AI-coding session histories tied to a campaign active from late September to early October. Its assessment that the actor was likely Chinese-speaking and financially motivated carries moderate confidence, not a confirmed attribution.
Files reveal an AI-assisted campaign
CrowdStrike says researchers examined open directories on infrastructure associated with the attacker. They found Claude Code session histories, ARTEX configuration files and Claude memory files, which the company says offered direct insight into the operator’s methods and tools. The campaign targeted South Korean financial organizations and resulted in data being taken, although CrowdStrike says the number of affected organizations remains unconfirmed.
“While this activity has not been attributed to a named adversary, the threat actor is likely a Chinese speaker and financially motivated,” writes Ashley Campion of CrowdStrike. The company says it bases that assessment on the use of ARTEX, a China-developed open-source agentic penetration-testing tool, and Chinese-language prompts seen in the activity.
The report describes two compromised systems at separate banks: a loan-progress inquiry service used by financial brokers and an employee mobile work-support system. CrowdStrike does not name the banks in its account. It says the targeted organizations overlap with those identified in industry reporting, while warning that the full scope is not established.
ARTEX connected to several language models
Analysis of the files points to a two-server setup. A Hong Kong-based address served as the main attacker-controlled infrastructure, while another server hosted the ARTEX instance CrowdStrike believes was used in the South Korean activity. The company says that instance used DeepSeek v4.1-flash as its primary language-model backend; the operator also used GLM-5.3 and Grok 4.6 in additional Claude Code sessions.
The tool’s role matters to the attribution question. CrowdStrike describes ARTEX as an agentic penetration-testing tool, not as a language model or proof of who operated it. The company’s report links observed use and Chinese-language prompts to its assessment, but says the activity has not been attributed to a named adversary. The evidence it describes establishes tool use more clearly than the operator’s identity.
Personal details do not establish identity
One Claude Code session asked for a security-researcher résumé containing personal details and claims about ARTEX-related work. The prompt gave an age of 26 but also initially supplied a birth date of September 22, 2007, a discrepancy that makes those details especially poor grounds for treating the résumé as verified identification. CrowdStrike says the information may belong to the person behind the activity but cannot definitively connect it to the threat actor.
The sessions also show the user asking Claude where stolen Korean data is commonly sold and requesting help finding Telegram groups that trade such information. Those questions support CrowdStrike’s view that the activity may have had a financial motive; they do not establish that the actor sold data or identify a buyer.
CrowdStrike’s report therefore presents two different levels of certainty: files and session histories link ARTEX and several language models to the campaign, while the person behind it and the total amount of data taken remain unresolved. The company’s published account names no confirmed adversary and does not give a final count of affected organizations.