The attackers probed the sector's side doors — loan-broker lookup pages, staff support systems and other services outside the main customer-facing platforms. They kept testing until several gave way.
At least six banks were targeted. Customer data leaked from Shinhan Bank, KB Kookmin Bank and Hana Bank, while BNK Busan Bank exposed information on 11 outsourced developers. Shinhan alone reported the loss of records approximately 25,000 customers.
NH NongHyup Bank and Woori Bank blocked the intrusions, according to financial authorities.
Authorities have identified 19 internet protocol addresses across 12 countries and territories linked to the attacks, which extended across major commercial banks, regional and online lenders, and brokerage houses.
Five addresses were located in the United States, while Japan, Sweden and Germany accounted for two each. The others were traced to Hong Kong, Singapore, Vietnam, Thailand, Malaysia, Spain, Latvia and South Korea. The locations identify infrastructure associated with the attacks, rather than necessarily revealing where the attackers were based.
The breadth and speed of the campaign have raised suspicions that hackers used AI to automate parts of their work. How much AI actually performed remains unconfirmed.
"There are two grounds for the suspicion. First, evidence emerged that an AI tool was likely installed on a server used in the attacks," said Yeom Heung-youl, a professor of information security at Soonchunhyang University.
Investigators found a Chinese-language console title associated with ARTEX, an open-source AI penetration-testing platform, on suspected attack infrastructure. Its presence points to possible use of the tool but does not establish precisely how it was deployed.
"Second, the attacks were carried out on multiple systems at the same time and on a wide scale," Yeom said.
The same attack IP addresses appeared at the four banks that reported breaches, according to the Financial Security Institute.
Breaches later surfaced at Hyundai Capital and two savings banks, Yegaram and Welcome. Those attacks used different IP addresses but followed a similar pattern, authorities said.
Lenders with missing or weak identity checks were breached, while those with stronger authentication and access controls held, he said.
That distinction puts the focus on existing security loopholes as much as the sophistication of the suspected tools. AI may accelerate the search for an opening, but the failures identified so far include basic weaknesses in verifying users and restricting access.
"Rather than a hacker acting alone, it looks like the hacker used AI agent technology," said Park Choon-sik, a professor of information security at Seoul Women's University.
ARTEX is designed to coordinate AI agents that gather information, identify vulnerabilities, plan possible intrusion routes and run security tests. Built for penetration testing, such software can also be misused to automate attacks. The language model, if any, used in the Korean breaches remains unidentified.
Regulators found a lookup service that displayed loan-application data without verifying identity and a staff system lacking controls over access from mobile devices. In another case, attackers exploited a known web vulnerability to install malware and steal log files containing customer information.
At Shinhan, an attacker entered randomized customer numbers into six loan-related services and attempted to bypass phone verification, according to incident reports obtained by Rep. Park Min-kyu.
Each blocked IP address was replaced with another, with the addresses spanning eight countries.
Areas that "drew relatively less management attention" became targets, Financial Services Commission Chairman Lee Eog-weon said on Oct. 4.
The incidents follow a succession of major breaches that exposed different weaknesses across Korea's digital economy.
Malware on SK Telecom's servers exposed USIM data associated with about 23 million subscribers in 2025. Illegal femtocells, or miniature mobile base stations, enabled criminals to siphon information from 22,227 KT users.
Coupang's exposure of 33.7 million accounts, disclosed in November 2025, was traced to a former employee who retained an authentication key.
The emerging concern is that AI could multiply the speed and scale at which such weaknesses are found and exploited.
"This is a completely new mode of attack. It is larger in scale and more advanced, and it could become more intelligent," Park said.
International cases illustrate that risk, although they do not establish what happened at the Korean institutions.
Check Point reported in September 2025 that criminals were discussing using HexStrike AI, a framework connecting AI models with more than 150 security tools, to exploit newly disclosed Citrix vulnerabilities.
OpenAI disclosed on Aug. 26 that research models operating with reduced safeguards during cybersecurity evaluations in July had bypassed isolation controls and compromised parts of its research infrastructure and Hugging Face's systems. The company said the models took dangerous actions that no human had directed.
That is "a very dangerous situation," Yeom said.
"In severe cases, they find a flaw and attack within a minute, so fast that defenders cannot respond," Park said.
For financial institutions, the immediate task is to establish where attackers entered, how far they moved and what access they retained.
"The priority is to confirm through log analysis the initial point of entry, any stolen accounts or privileges, movement inside the network and the actual scope of data taken out," said Hwang Suk-jin, a professor at Dongguk University's Graduate School of International Affairs and Information Security.
"More than the size of the leak, how long the attacker stayed inside and what level of privileges it secured can be an important yardstick," Hwang said.
The longer-term challenge is to shorten the time between discovering a weakness and closing it.
"We need the technology and the system to quickly find vulnerabilities in every internet-facing server using AI, and to patch them within a short period," Yeom said.
That requires close cooperation with software suppliers, automated patch testing and decision-making procedures that allow institutions to act quickly, he said.
Park called for mandatory adoption of zero trust, an approach that verifies users, devices and access requests rather than assuming activity inside a network is safe.
"An attacker succeeds if one in 100 attempts works. A defender is breached if even one in 100 gets through," Park said. "Security is hard to achieve without a little inconvenience."
He likened passwords to seatbelts: drivers still need to wear them even when their cars have brakes. Yet after one breach after another, many users continue to rely on the same password everywhere.
AJP Takeaways
- Hackers breached internet-facing support systems at Shinhan Bank, KB Kookmin Bank, Hana Bank and BNK Busan Bank in late September 2026, exposing data on 25,727 Shinhan customers, while NH NongHyup Bank and Woori Bank blocked similar attacks.
- South Korea's Financial Security Institute found traces of ARTEX, an open-source AI penetration testing tool on GitHub, on attack servers, and experts pointed to the simultaneous, wide-scale pattern as a sign of AI-driven automation.
- Regulators found lookup services without identity checks and missing access controls, and lenders with strong authentication held off the attacks, according to Soonchunhyang University professor Yeom Heung-youl.
- Experts called for AI-driven vulnerability scanning and rapid patching, mandatory zero trust security and multi-factor authentication for individuals, as AI agents cut the time from flaw discovery to attack to minutes.
Copyright ⓒ Aju Press All rights reserved.