Korea's Banks Were Shown the AI Attack Two Weeks Before It Hit Them

South Korea's financial security agency showed bank security staff an AI-agent attack on a domestic lender on 17 September. Within two weeks, a wider campaign using the same kind of tooling hit seven financial firms. Documents the Financial Security Institute (FSI) submitted to lawmaker Han Chang-min, disclosed on 11 October, show how it unfolded.

The warning

At a seminar on financial AI security threats attended by about 160 security officers, the FSI walked through a case involving a firm it called Company A. The attacker told an agent built on the Chinese model DeepSeek to scan the financial sector for servers running Oracle WebLogic. The next day the agent exploited a remote command execution flaw on one company's server and planted a web shell. Company A caught the intruder just before more malicious code could be pulled onto internal servers.

The FSI's advice was direct: assessments done once or twice a year will not catch this, and internet-facing systems need constant measurement and verification. Staff from four of the firms that were later breached were in the room.

The breaches

The attacks started in late September. Shinhan Bank reported about 25,000 affected customers on 1 October, with names, phone numbers, annual income and loan limits exposed. Yegaram Savings Bank lost data on about 40,000 customers. KB Kookmin, Hana, BNK Busan, Welcome Savings Bank and Hyundai Capital also confirmed leaks. Woori and NH NongHyup said they blocked the attempts.

CrowdStrike said in a 7 October report that the attacker used ARTEX, an open-source penetration-testing tool developed in China that hands much of its work to AI agents. The copy it examined ran on DeepSeek V4.1-Flash, and the attacker used Anthropic's Claude Code and other models in separate sessions. CrowdStrike assessed with moderate confidence that the attacker is a financially motivated Chinese speaker, not linked to a known group. This time the targets were services the banks ran themselves, including a broker loan-inquiry page and an employee mobile app.

ARTEX's developer said on 8 October that the tool will no longer be updated and will go closed source because of misuse. Copies already downloaded keep working.

The cost

The Korea Times reports that demand deposits at the five largest lenders fell from 675.5 trillion won on 30 September to 655.4 trillion won, a drop of more than 20 trillion won (about $15 billion) in 11 days. KB Kookmin plans to lift its security budget above 100 billion won next year, from 86.1 billion won this year. Woori is weighing a rise of more than 20%. Several banks are now using AI to attack their own systems first, including NH NongHyup's planned AI red-teaming system. The heads of the five largest banks testify at a National Assembly audit on 19 October.

What this means for Indian BFSI

I wrote about the first wave of these attacks on 4 October. What's new is that the warning came first, and it wasn't enough. The attack itself was not sophisticated: scan for a known product, exploit a known flaw, drop a web shell. The difference is the speed and scale an agent brings to that routine. Annual VAPT cycles were built for human attackers working at human pace. Indian banks and NBFCs with large internet-facing estates, broker portals and staff apps should be asking one question this week: how quickly would we know if an agent found one unpatched server?

Source: AJP: Korean banks had AI hacking warning weeks before data breaches