One man, working alone from southern China, broke into five South Korean banks in two weeks. He brought no crew and no custom malware. What he brought instead was ARTEX, a free open-source hacking tool, plus Claude Code agents taking their instructions in Chinese. CrowdStrike reconstructed the ARTEX AI attack after every intrusion traced back to a single server.
The campaign ran from late September into early October, and the fallout reached ordinary account holders within days. South Korea’s financial regulator issued a consumer alert on October 6, warning affected customers about phishing attempts and loan scams. Several banks saw services go dark while their teams scrambled.
How the ARTEX AI attack unfolded
ARTEX arrived as an open-source project built in China, and it does something older scanners never could. Rather than grind through a fixed checklist, it lets a language model decide what to probe next. The operator pointed it at Korean bank systems and let it hunt.
DeepSeek v4.1-flash served as the engine behind the ARTEX AI attack. Investigators believe he reached it through a reseller proxy at xcai[.]pro. He added GLM-5.3 from Zhipu AI and Grok 4.6 across extra Claude Code sessions. Spreading the work across providers meant no single account carried the whole campaign.
A human stayed in the loop throughout the ARTEX AI attack. He wrote the prompts, read the output, and decided what was worth taking. The tooling collapsed weeks of scouting into hours, but nothing moved without his say-so.
What the banks lost
The ARTEX AI attack caught at least five lenders. Reported breaches name Shinhan Bank, KB Kookmin Bank, Hana Bank, Yegaram Savings Bank, and BNK Busan Bank. Shinhan put its own exposure near 25,000 people, while Yegaram Savings Bank reported a figure closer to 40,000. CrowdStrike has not confirmed a final victim count, so those numbers may still climb.
The stolen records held personal details and credit card data. At one bank the operator broke into a loan progress inquiry service used by financial brokers. At another he compromised a mobile work-support system built for staff. Neither of those sits at the front of a bank’s defenses, which is part of why they fell.
The forgotten edges of a network
Security budgets go to the obvious targets: core banking platforms, payment rails, the customer app. Broker portals and internal staff tools rarely attract the same scrutiny, though they often reach into the same databases. An ARTEX-style attack has no sense of which system matters, so it works through everything exposed until something gives.
The résumé that cracked the case
His own security failed in a thoroughly ordinary way. He left directories on his own servers open to the public internet, and researchers walked straight in. Those folders held Claude Code session logs, ARTEX setup files, and Claude memory files. Together they read like a running diary of the operation.
One session went further than the rest. He had used the same tooling to write his own résumé. That document carried ID details, contact data, and the Telegram handle YY520CN. A separate prompt placed him in Maoming, Guangdong.
The résumé described a 26-year-old who had studied at the South China University of Technology. But a birth year of 2007 contradicts that age, so the profile stays unreliable. CrowdStrike reads the operator as a Chinese speaker chasing money rather than a state operative.
Selling the data looked like an afterthought. Once he had it in hand, he asked Claude to suggest Korean Telegram groups that trade stolen records. That request sat in the very logs that later exposed him.
A lower bar for everyone
Hitting five banks once demanded a crew, patience, and years of practice. With a model supplying the technical judgment, one amateur covered the same ground in a fortnight. The ARTEX AI attack produced five victims, tens of thousands of exposed customers, and a trail of wide-open logs.
News of the ARTEX AI attack pushed its developers to close the source and stop updates. That came too late to matter much. Copies of the original code already circulate as English and Korean forks, so the capability is out for good.
Protecting yourself after a bank breach
Customers of the banks hit by the ARTEX AI attack now face a predictable second wave. Criminals holding real names, phone numbers, and partial card details write convincing phishing messages, because the details check out. The regulator’s warning about loan scams points at exactly that pattern.
Simple habits help more than panic here. Watch your statements, treat any surprise call as suspicious, and never confirm details to someone who contacted you first. Banks do not ask for passwords or full card numbers over the phone.
A VPN will not undo a breach inside a bank’s own systems, and nobody should pretend otherwise. What it does do is limit what you give away everywhere else. It hides your traffic from anyone watching your network and blocks the trackers behind targeted phishing. NordVPN, Surfshark, and Windscribe all bundle threat protection that flags known phishing domains before a page loads.
Final Thoughts
AI tooling has quietly lowered the bar for attackers, and Korean banking just absorbed the proof. The operator behind the ARTEX AI attack left his own résumé sitting on an open server. He still walked away with records on tens of thousands of customers. Skill mattered far less than access to the right tools.
Defenders face the same math in reverse, because the systems that fall first are the forgotten ones. The broker portal, the staff utility, the service nobody has audited since launch. Banks that inventory those assets now will fare better than the ones waiting for a notification letter. For everyone else the advice stays dull but useful: watch your accounts, distrust cold contact, and limit what you leak.
