South Korea's president, Lee Jae Myung, has ordered a full investigation after a run of data thefts hit several of the country's banks and lenders, and one detail in the early evidence has made security people sit up. GBHackers reports that investigators found a Chinese-language string reading "AI autonomous penetration testing console" in infrastructure linked to the Shinhan Bank breach. It matches ARTEX AI, an open-source framework that automates reconnaissance, vulnerability discovery and attack planning.
What is actually known
Shinhan Bank has about 25,000 customers affected, Yegaram Savings Bank around 40,000, and Hyundai Capital 146 housing-loan agents. KB Kookmin, Hana and BNK Busan Bank have each reported separate incidents. The exposed fields were names, phone numbers, annual income and loan limits, with resident registration numbers in some cases. No payment credentials or fraudulent transactions have been confirmed so far.
The authorities have been careful, and so should we. A string in some infrastructure does not prove ARTEX AI was used. The tool is public, so anyone can run it, and attribution is still open. The more concrete evidence points at credential stuffing, which is replaying passwords from older leaks against login pages.
Why the AI angle is less exciting than it sounds
If automation is stuffing credentials into a bank login at scale, the tooling is not the interesting part. The interesting part is that the defences which stop it were all available before anyone wrote an LLM wrapper: phishing-resistant multi-factor authentication, rate limits that bite per credential rather than per IP, breached password screening, and anomaly detection on sessions. The Financial Services Commission and Financial Supervisory Service have now called emergency meetings and are pressing banks for exactly that: security inspections and phishing-resistant MFA.
Where AI does change the picture is pace and patience. An agent will try a thousand slow, varied approaches without getting bored, which makes the cheap signals (identical user agents, bursts from one network) less reliable. Defenders need signals that are expensive to fake, like device binding and behavioural patterns across a whole session.
Questions for your own login estate
- What share of customer and partner accounts can still sign in with a password alone? That is your exposure to this exact attack.
- Do you throttle by account as well as by source address? Distributed attempts defeat per-IP limits.
- Are you screening new and existing passwords against known breach corpora?
- Would you notice a quiet, low-volume campaign running for three weeks? Check what your alerting says about a 2 percent failure rate that never spikes.
- Do agents or brokers with access to income and loan data (the Hyundai Capital pattern) have the same MFA as staff?
For help reviewing authentication and abuse detection on a customer-facing platform, email sales@halfteck.com.