Hackers used ARTEX, an AI-driven penetration testing framework, in a campaign targeting South Korean financial organizations between late September and early October 2026.
Public reports linked seven institutions to the incidents, with thousands of customers affected. However, CrowdStrike said the total number of affected organizations remained unconfirmed in its October 7 report.
The supplied Hunt.io investigation places the reported impact at approximately 66,000 to 67,000 individuals and around 2,200 corporations.
Exposed information reportedly included personal details, employment records, income data, and loan information. These figures remain reported estimates rather than independently reconciled totals.
CrowdStrike assessed with moderate confidence that the operator was Chinese-speaking and financially motivated.
Investigators found AI sessions asking where stolen Korean data could be sold, including requests for help locating Telegram sales groups. The campaign has not been attributed to a named threat group.
ARTEX Hackers Target Banks
The investigation centered on 38.244.50[.]120, a server hosting an ARTEX instance. Exposed directories contained Claude Code session histories, configuration files, and memory documents that revealed the operator’s activities. Such directories allow files to be accessed publicly without authentication.
According to the supplied Hunt.io analysis, researchers identified another exposed server at 23.248.226[.]250 in Hong Kong. Its directory contained 42 files across two subdirectories, totaling approximately 105 MB.
A recovered credentials file referenced the ARTEX server, while shell scripts supported reverse SSH tunnels, service relays, and memory synchronization. These artifacts provided stronger links between the two systems than matching network indicators alone.

The AI setup combined Claude Code with alternative model providers rather than relying solely on Anthropic services. CrowdStrike identified DeepSeek v4.1-flash as the primary backend, supplemented by GLM-5.3 and Grok 4.6. Requests likely passed through the reseller domain xcai[.]pro.
Hunt.io’s supplied findings describe custom routing scripts that handled model discovery and added reasoning settings to outgoing requests.
Researchers also recovered 19 reusable agent skills covering reconnaissance, source review, authentication testing, payment analysis, and Android application inspection.
Additional files included proxy management scripts, reconnaissance utilities, and a staged Java Naming and Directory Interface exploitation setup. The latter used LDAP responses and DNS callbacks to test whether remote commands could execute.
A proxy service at 154.201.79[.]246:443 reportedly produced 79 unique exit addresses from 100 tested identities. CrowdStrike separately listed that IP among proxies used during the campaign.
Researchers also recovered cryptocurrency-related scripts, Linux privilege-escalation binaries, and mobile-device tools.
Their presence shows a wider operational toolkit, but does not establish that every artifact was used against Korean institutions, hunt.io said.
Indicators of Compromise
| Type | Indicator | Context |
|---|---|---|
| IP | 38.244.50[.]120 | ARTEX console and jump host (Los Angeles, AS174 Cogent); ports 443, 8090, 8000, 5918, 18888 |
| IP | 23.248.226[.]250 |
Note: IP addresses and domains are intentionally defanged (e.g., [.]) to prevent accidental resolution or hyperlinking. Re-fang only within controlled threat intelligence platforms such as MISP, VirusTotal, or your SIEM.
Join 16,000+ SOC teams using ANY.RUN to streamline threat investigations and reduce manual effort. Explore for your team