One Man, One AI Stack, Nine Banks: CrowdStrike Details the AI-Driven Korean Bank Hacks

What Happened

The attacks ran from late September to early October 2026. At least nine South Korean banks have "disclosed attacks or been reported as attacked" (Reuters' wording), with Shinhan Bank, KB Kookmin Bank and Hana Bank named. Shinhan admitted last week that about 25,000 customers' personal information was leaked — names, contact details, income, credit limits; KB Kookmin confirmed 119 customers affected. What was breached was not the core ledger but "business systems": one bank's loan-progress inquiry service (used daily by financial brokers) and another bank's employee mobile work-support system. CrowdStrike's evidence chain came from what the attacker left on public servers: a Hong Kong IP as primary infrastructure, an ARTEX instance running at 38.244.50[.]120, and open directories holding Claude Code session logs, ARTEX config files and a Chinese-language pentesting prompt (.claude/CLAUDE.md).

The stack: ARTEX is an open-source AI penetration-testing agent published on GitHub this July by a Chinese security engineer — it plugs into external LLMs to find vulnerabilities automatically, with the author stating it's for learning and research only, not real-world use. The attacker ran DeepSeek v4.1-flash as the primary engine (via a suspected API proxy, xcai.pro), with GLM-5.3 (Zhipu) and Grok 4.6 assisting inside Claude Code sessions. The leaked sessions show the attacker asking Claude where Korean data is usually sold and getting help finding Korean Telegram data-trading groups, and having Claude draft a security-researcher resume as a "26-year-old, South China University of Technology, Maoming, Guangdong." CrowdStrike says the attacker's identity is unconfirmed, with no attribution to any organization; Reuters called the phone number in the report and the man who answered said he knew nothing about it.

Key Facts

1. Attacker identity unconfirmed: CrowdStrike's actual words: no attribution to a named adversary; the actor is assessed as "likely a Chinese speaker and financially motivated" — moderate confidence, based on the Chinese-developed ARTEX plus observed Chinese-language prompts. The "26-year-old Guangdong resume" is what the attacker wrote into his own leaked sessions — unverified. The attacker's identity and the full scope of stolen data are both unconfirmed.
2. Business systems were breached: What was breached were business systems — loan inquiries, employee mobile work — not the core ledger. The attacker's control server had open directories exposing Claude Code sessions, config files and the Chinese prompt; CrowdStrike reconstructed the full attack chain from them.
3. The report describes a person using an AI toolchain: The report describes one person using an AI toolchain to automate and accelerate pentesting — not an autonomous agent planning attacks on its own. ARTEX is open source and downloadable from GitHub. Anthropic recorded days ago that GLM-5.3 writes exploits at near the level of its frontier model Mythos Preview.
4. How the numbers break down: 25,000 is Shinhan-confirmed; 119 is KB-confirmed; the full scale of stolen data is unconfirmed per CrowdStrike; "nine" is "disclosed or reported" — how many were actually breached and how much was taken is under police investigation. The financial regulator has held an emergency meeting and President Lee Jae Myung has demanded a thorough investigation.

Context

The timeline: starting late September, South Korean banks began disclosing attacks; last week Shinhan admitted ~25,000 customers' data leaked; on October 7 CrowdStrike published the full technical report; this week South Korean police opened a case, the financial regulator held an emergency meeting, and President Lee Jae Myung demanded a thorough investigation and robust response. ARTEX was published on GitHub in July; in March, Anthropic's frontier model Mythos prompted a debate over whether to release it because of its offensive cyber capabilities.