AI-Assisted Attacks on South Korean Banks: What Actually Happened
Min-jun is a loan broker in Busan. Every morning he logs into his bank’s broker portal to check where his clients’ loan applications stand. It is a boring website. Nobody has ever thought of it as a target.
In late September, that kind of website became the way into one of Korea’s biggest banks.
Between September 27 and 30, attackers broke into at least seven South Korean banks and lenders. They used a free AI tool that hunts for weak spots on its own. The President of South Korea went on live TV and said AI now makes it possible to hack with ease, even without special skills.
So what really happened? And how much of the panic is justified?
South Korea Bank Attacks: Key Facts
- Dates: Attacks ran from September 27 to early October 2026
- Banks hit: Shinhan Bank, KB Kookmin Bank, Hana Bank, BNK Busan Bank, Yegaram Savings Bank, Welcome Savings Bank and Hyundai Capital
- People affected: Roughly 65,000 to 68,000, depending on who is counting
- Biggest leaks: Yegaram Savings Bank (about 40,000 people) and Shinhan Bank (25,727 people)
- Smaller leaks: KB Kookmin (119 people) and Hana Bank (89 people)
- Data taken: Names, phone numbers, birth dates, yearly income and loan limits
- Not taken: Passwords and one-time login codes. Regulators have confirmed no customer has lost money.
- The AI tool: ARTEX, a free tool built in China to test company security
- Who is behind it: Not officially named. CrowdStrike says a likely Chinese-speaking person motivated by money.
- Also hit, no leak confirmed: Woori Bank and NH NongHyup Bank saw break-in attempts
The attackers did not go after the vault. They went after the side doors.
At Shinhan, they got into a loan progress service used by outside brokers like Min-jun. At KB Kookmin, they got into a mobile system used by bank staff. At Hana, it was a sales support system.
None of these hold customer money. But all of them hold customer details, and all of them are easier to break into than the core banking system. Authorities say the attacker did not aim at one bank. They hit many at once and got through the ones with weaker defences.
That is the real lesson. Your strongest wall does not matter if there is an open window around the back.
What ARTEX Does
ARTEX is a tool that uses AI to do what a hired security tester does. It scans a company’s systems, finds weak points, plans a way in, tries it, and checks whether it worked. All with very little human input.
It was made for authorised testing. It was released free for anyone to download. And the latest version came out on September 24. Three days later, the attacks on KB Kookmin began.
One detail matters here. ARTEX is not an AI by itself. It plugs into other AI services to do its thinking. In these attacks, CrowdStrike says it ran mainly on DeepSeek. The attacker also used other AI tools, including Anthropic’s Claude Code, to help with extra work.
The Attacker Left the Door Open
Here is the part that is almost funny.
The attacker left folders open on their own servers. Inside were configuration files and full records of their AI chats. CrowdStrike found them and could see exactly how the campaign was run.
The records showed something else. The attacker had no clear plan to sell the stolen data. They asked the AI to suggest Telegram groups where Korean data gets traded. They also asked it to write a security researcher résumé listing the bank attacks as achievements.
That résumé held personal details. CrowdStrike says they probably belong to the attacker, but it cannot prove it. One detail did not add up. The age given was 26, yet the same person first typed a birth date in 2007. We are not repeating the details here.
CrowdStrike says it does not know who this is. Korean police have put 28 investigators on the case. Authorities plan to ask other countries for help, since the attacks were routed through overseas computers.
Did AI Hack the Banks by Itself?
No. A Korean official put it plainly: a hacker used AI as a tool. The AI did not act without human involvement.
But AI did change two things.
Speed. One person ran attacks on many companies in about two weeks. CrowdStrike says this shows how AI lets a money-driven attacker break into several places in a short time.
Skill. Work that once needed a team of experts can now be handed to a tool. Some people in Korea have raised the possibility that this was a lone individual, not a professional gang.
We should be careful here. The AI link is strongly suggested by the evidence, but Korean authorities have not officially confirmed which tools were used. Anthropic had not commented publicly in the reports we reviewed.
Banks Were Slow to Notice
This number should worry you more than the AI part.
Shinhan took 15 hours and 26 minutes to detect its breach. Hana took 41 hours and 44 minutes. KB Kookmin took 67 hours and 41 minutes. That is nearly three days.
An automated tool does not wait around for three days. Regulators named weak login checks and too much stored data as the main problems. They warned that similar failures will face stern penalties.
What Korea Is Doing
Financial regulators called an emergency meeting on the October 4 holiday and pulled the heads of the biggest banks in. Police opened a formal investigation. Banks are working out compensation plans.
Regulators are also thinking about relaxing a rule that keeps bank networks separate from the outside. The idea is to let banks use outside AI security services to find weak spots faster. That is a real tradeoff. More connection means more exposure.
After the news broke, the ARTEX developer pulled the project and made it closed-source. He said it was built for authorised testing. Copies already exist, and people have made English and Korean versions.
Indicators of Compromise (IOCs)
These come from CrowdStrike’s report. Block them and search your logs for them.
# AI-assisted attacks on South Korean banks (ARTEX campaign)
Attacker server (hosted the ARTEX tool):
38.244.50[.]120
Proxy addresses used during the attacks:
101.53.80[.]20
205.214.59[.]31
124.155.252[.]63
154.201.79[.]246
23.248.249[.]90
23.158.220[.]98
103.248.148[.]84
203.160.133[.]172
209.209.85[.]38
AI access service the attacker likely used:
xcai[.]pro
# Note: attackers change addresses fast.
# Treat this list as a starting point, not a full defence.What To Do Now
Priority 1: Find your side doors. List every partner portal, staff mobile app and sales tool that touches customer data. These were the way in. Check who can reach them from the internet.
Priority 2: Fix weak logins. Regulators named this first. Turn on a second login check everywhere, especially for outside partners.
Priority 3: Delete data you do not need. Shinhan lost income and loan limits. Ask why a broker portal holds that much. What you do not store cannot be stolen.
Priority 4: Get faster at noticing. Three days is too slow. Set alerts for strange logins at odd hours and unusual bulk downloads.
Priority 5: Test yourself the same way. Attackers will use tools like ARTEX. You can run similar tools against your own systems first, with permission.
Priority 6: Check your logs for the addresses above. Look back to late September.
The ZyberWalls Perspective
The scariest part of this story is not the AI. It is how ordinary the targets were.
A loan broker portal. A staff phone app. A sales system. Nobody guarded them like the vault, so they became the way in. AI did not invent that weakness. It just found it faster than any human could.
The other uncomfortable fact is who did this. Possibly one person. Possibly someone with no plan to even sell the data, who then asked an AI to turn the crime into a résumé. That is the new picture. The skill needed to hurt a large bank is dropping, while the speed of attack is rising.
Min-jun is still logging into his portal every morning. His clients’ names, phone numbers and incomes may already be in someone’s folder. Banks say no money was lost. That is true today. But leaked details have a long life, and scammers love a victim whose income and loan limit they already know.
Defenders now face attackers who work at machine speed. Waiting three days to notice will not survive that.
Stay Alert. Stay Human. Stay Safe.
— ZyberWalls Research Team