ARTEX Developer Closes AI Pentest Tool After Korea Bank Hacks
TL;DR
- One operator ran ARTEX with four LLM backends simultaneously, compressing the full attack cycle from reconnaissance to exfiltration into hours per target.
- Attribution came from the attacker's own exposed directories containing Claude Code session histories and Chinese-language prompt files, not from any bank's network monitoring.
- Two banks on identical architecture repelled the same ARTEX attacks, confirming the breach was a function of peripheral portal hardening gaps, not tool sophistication.
CrowdStrike has traced a late-September string of intrusions at Shinhan Bank, Yegaram Savings Bank and other South Korean financial firms to an operator running ARTEX, an open-source agentic penetration-testing tool built in China. Its developer, who goes by Autumn-27 on GitHub, responded on October 8 by converting the project to closed source and halting further releases.
"The ARTEX project will no longer be updated and will be converted to a closed source. There will be no release of any version or maintenance support in the future," Autumn-27 wrote. The developer said ARTEX "was originally designed for the purpose of learning and research" and that the attacks "had nothing to do with them."
ARTEX runs DeepSeek v4.1-flash as its primary backend, with Z.ai's GLM-5.3 and SpaceXAI's Grok 4.6 as supplementary models, and the operator reportedly reached those APIs through a reseller at xcai[.]pro, The Hacker News reports. CrowdStrike Intelligence identified two servers behind the campaign: a Hong Kong IP that hosted the backbone, and a second machine at 38.244.50[.]120 running the ARTEX instance. Investigators found exposed Claude Code session histories, Claude memory files and ARTEX configuration files on the Hong Kong server.
Those Claude artifacts also surface the operator's downstream plans. CrowdStrike reports that "the threat actor asked Claude where threat actors typically sell Korean data breach information and asked Claude for assistance in finding Korean Telegram data sales groups," with the Telegram handle "@YY520CN" referenced in the prompts. The actor is suspected to be Chinese-speaking and financially motivated and is not currently tied to any tracked group. It's the latest AI-weaponization alert on our cybersecurity tracker, where Japan's SoftBank and Daiwa breach disclosures landed earlier the same week.
South Korea's Financial Services Commission and Financial Supervisory Service issued a public warning telling consumers to "exercise vigilance and be on the lookout for potential phishing attacks and loan scams stemming from leaked data."
What others are reporting
-
Infosecurity Magazine Read →
Traces the OPSEC failures enabling attribution: linked IPs, leaked Telegram handle YY520CN, Maoming geolocation, and Chinese-language Claude Code prompt files left in open directories.
AI tooling can enable a financially motivated threat actor to conduct multiple intrusions within a short time span.
-
Cyber Security News Read →
Documents the attacker querying Claude about where to sell stolen Korean data and how to find Telegram data-sale groups, confirming financial motivation and AI-assisted post-exploitation.
AI can speed up target research, scan analysis, testing, scripting, record keeping, and the use of existing tools.
-
MLQ News Read →
Names CrowdStrike as the attributing party and reports South Korea's FSC is actively weighing a reversal of network-separation rules triggered by the campaign.
The operator used DeepSeek V4.1-Flash, Zhipu AI's GLM-5.3, Grok 4.6 and Anthropic's Claude Code during the activity.
-
Cyber Press Read →
Details how exposed ARTEX config and memory files in open directories gave researchers a full map of the attacker's prompts, infrastructure, and target selection methodology.
The findings show how an attacker combined agentic AI with conventional offensive capabilities to support operations against several financial-sector targets.
-
TechTimes Read →
Provides 30-43 hour undetected dwell times, an IP breakdown across 359 addresses (65.7% US, 14.8% China), and notes two banks repelled identical attacks, quantifying the defensive gap.
Hackers have begun using AI agents, so the frequency and scale of attacks are bound to grow more severe. - Kim Myeong-ju, Barun AI Research Center
Originally reported by thehackernews.com
Read the original article →Original headline: ARTEX Developer Takes AI Pentest Tool Closed-Source After Attribution to South Korean Bank Hacks