A malicious actor leveraged the ARTEX penetration-testing tool, combined with several generative AI models, to infiltrate South Korean banking systems. The incident highlights the growing sophistication of automated cyberattacks driven by financial gain.
Certaines infrastructures financières sud-coréennes ont subi une série de violations de données entre la fin septembre et le début octobre 2026. Pour ce faire, l'attaquant a compromis un service de suivi de prêts pour courtiers et un système mobile de support aux employés. L'analyse de l'équipe CrowdStrike Intelligence montre que surtout l'utilisation de l'intelligence artificielle agentique pour piloter les intrusions.
The cybercriminal relied on ARTEX, an open-source penetration-testing tool developed in China. The platform acts as an orchestration engine, automating vulnerability discovery and exploitation by querying multiple large language models (LLMs). The ARTEX instance involved in the attacks primarily used DeepSeek V4.1-Flash, supplemented by Zhipu AI’s GLM-5.3 and Grok 4.6 through Claude Code sessions.
“The use of agentic AI tools alongside traditional offensive capabilities highlights the evolution of adversarial tactics,” CrowdStrike Intelligence stated.
Configuration Errors
However, the malicious actor made several operational mistakes. A server based in Hong Kong, identified as the campaign’s primary infrastructure, exposed publicly accessible directories.
Analysis of these exposed servers enabled investigators to recover traces of the campaign.
According to CrowdStrike, these traces also point to a financially motivated operation. The attacker reportedly queried an AI model to identify Telegram channels specializing in the resale of South Korean databases.
Profiling the Hacker
The attacker’s conversation histories with LLMs also revealed the generation of a cybersecurity researcher’s résumé.
The document described the results of the ARTEX attacks and included personal details, claiming the individual was 26 years old, based in Guangdong, China, and used the Telegram handle @YY520CN. The same identifier had also been observed in vulnerability research activities targeting an NFT marketplace and a Chinese payment platform.
Beyond using a résumé to showcase these exploits, CrowdStrike’s analysis highlights how integrating LLMs into cybercriminal workflows can significantly reduce the time required to execute offensive campaigns.

Comments