According to CrowdStrike, one attacker, one AI-agent pentesting tool and several language models were enough to breach South Korean bank systems. The report is valuable not only for the attack but for what it exposed: Claude Code session histories left open on the attacker’s own server.
CrowdStrike published its analysis on Wednesday, October 8, 2026 (reported by Korea JoongAng Daily), and BleepingComputer expanded on it on October 10. The firm assesses, with moderate confidence, that the actor is likely Chinese-speaking and financially motivated. The attacks took place between late September and early October.
The banks affected by the series of breaches are Hana Bank, KB Kookmin Bank and Shinhan Bank, and financial authorities have opened investigations. Among the compromised systems, according to South Korean press, are a loan inquiry service used by brokers and a mobile employee-support system at another bank. BleepingComputer reports theft of customers’ personal and card data, and system outages at some institutions.
ARTEX is an open-source agentic pentesting suite developed in China: instead of a person running each tool, an LLM directs the penetration tests. According to CrowdStrike, the attacker used DeepSeek v4.1-flash as the main engine, with support from GLM-5.3 (Zhipu AI) and Grok 4.6 in some sessions, and Claude Code sessions as the working harness. BleepingComputer cites that DeepSeek access reportedly went through an API proxy or reseller (xcai[.]pro).
Researchers found open directories containing Claude Code session histories, ARTEX configuration files and Claude memory files. In one session, the attacker asked for a résumé to be drafted for a 26-year-old with training from the South China University of Technology in Guangdong. But CrowdStrike cautioned that this personal data is not reliable enough for a firm identification.
The technical barrier to attacking at scale has dropped: a single operator can orchestrate reconnaissance and exploitation with agents and cheap models. For a company in Guatemala or LATAM, and especially in finance, this means (1) speeding up patching of internet-facing services, because automation shortens the time between discovery and exploitation; (2) watching low-visibility internal tools, such as broker portals or employee-support apps; (3) applying the inverse lesson to your own agents: an agent’s session histories and memory are sensitive data, and no directory should be left open. Want to assess your attack surface? info@guatemalia.com.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel