Cyberattacks on South Korean banks have mounted since late September. Shinhan Bank, KB Kookmin Bank, Hana Bank, and BNK Busan Bank confirmed personal data leaks, while two other banks were attacked without suffering a breach. South Korea's Financial Services Commission called an emergency meeting on October 2, describing the situation as highly unusual. The Korea Financial Security Institute identified the tool used: ARTEX AI, an open-source autonomous penetration testing system based on large language models and distributed on GitHub.
Key points
- What happened: Personal data belonging to around 25,000 people at Shinhan Bank, 119 at KB Kookmin Bank, 89 at Hana Bank, and 11 at BNK Busan Bank was exposed. South Korea's FSC held an emergency meeting, calling the situation "highly unusual."
- Why the attacks worked: All targeted internal systems used by employees or partners rather than customer-facing services, where authentication and access controls were relatively weak.
- What was used: The Korea Financial Security Institute identified the open-source LLM-based autonomous penetration testing tool ARTEX AI. Humans operated it, while AI automated reconnaissance, vulnerability discovery, and attack execution.
- An asymmetric problem: Attackers need a GitHub account and an LLM API key. Defenders must protect countless internal systems used by employees every day.
The chain began with Shinhan Bank
The first breach occurred at Shinhan Bank between September 29 and 30 local time. Information belonging to approximately 25,000 people was exposed. Attackers bypassed authentication in a mobile portal for loan brokerage agents, obtaining names and telephone numbers as well as 66 resident registration numbers and 97 connected identification (CI) records. Bank president Jung Sang-hyuk apologized and pledged full compensation for victims.
Ordinary consumer-facing banking services were unaffected. Attackers targeted internal business systems used by employees or contractors, a pattern shared by all the affected banks.
At KB Kookmin Bank, an employee mobile business support system leaked information on 119 people, including names, telephone numbers, addresses, and resident registration numbers. Hana Bank exposed 89 records through a sales support system, and BNK Busan Bank exposed information belonging to 11 outsourced workers.
Woori Bank and NH NongHyup Bank were also targeted, but the particular vulnerabilities the attackers sought did not exist in those systems, so the attempts caused no reported harm.
After the Financial Security Institute shared attack-source IP addresses identified in the Shinhan Bank breach across the financial sector, other banks inspected their access logs and discovered further incidents. According to Herald Business, the attacks extended beyond commercial banks: logs at two savings banks are also being analyzed for similar activity. Yegaram Savings Bank has already disclosed a data leak on its website, suggesting the confirmed victim count may rise.
What is ARTEX AI?
The investigation by the Financial Security Institute identified traces of ARTEX AI. South Korean reports say the attacking server's HTML page title contained the Chinese text "AI 自主渗透测试控制台" (AI autonomous penetration testing console).
ARTEX AI is an open-source penetration testing system available on GitHub. Using LLMs, it can autonomously collect information, identify vulnerabilities, plan attack paths, run security tools, and verify results. It was released this year as a winning project in an "Agent+" attack-and-defense challenge hosted by Baidu's Security Response Center and is aimed primarily at Chinese-speaking users.
Joho no Todai found that its GitHub repository explicitly describes the project as an "AI 自主渗透测试系统" (AI autonomous penetration testing system) and that it also supports English and Korean.
Penetration testing tools are designed for legitimate security assessments and are lawful tools in themselves. The problem is that attackers repurposed their automation capabilities.
A Financial Security Institute official interviewed by Herald Business said the attack data contained signatures specific to ARTEX, making the tool identifiable. The official also emphasized an important distinction:
AI was indeed used in the attacks, but it did not act independently without human involvement. A hacker used AI as a tool.
The institute stressed that AI did not spontaneously attack banks; a person used it as a weapon. But that distinction offers limited reassurance. If AI can automate reconnaissance, vulnerability identification, and exploitation, the cost and time required for attacks fall sharply. In August, 21 Taiwanese government systems were compromised by an open-source AI agent (in Japanese).
Customer-facing defenses strong, internal systems weak
The attacks exposed a structural blind spot in South Korean financial institutions.
Every attack targeted internal business systems used by employees or contractors. Shinhan's loan broker portal was compromised through an authentication flaw; KB Kookmin's employee mobile system and Hana's sales support system were also targeted. Customer-facing online and mobile banking systems were not attacked.
The Financial Security Institute official noted that security for consumer electronic banking services had been strengthened substantially, while internal staff systems were managed less rigorously. Financial institutions did not fully understand where those internal vulnerabilities remained.
The reason Woori Bank and NH NongHyup Bank avoided losses fits the same pattern. The specific weaknesses the attackers targeted simply were not present in those two systems. That is different from detecting and repelling the attacks.
The attacks have not ended
According to the Financial Security Institute, attackers used multiple IP addresses, with two or three overlapping across the banks. Some source IPs used against Shinhan and Kookmin were identical, suggesting one attacker or a coordinated group. When one IP was blocked, the attackers switched to another and continued. The institute described blocking IPs as "only a stopgap." The attacks were continuing at the time of writing.
The risk of money being stolen directly from accounts is considered low. Attackers entered systems to view and retrieve information. But exposed names, telephone numbers, resident registration numbers, and credit information could be used for voice phishing, or telephone fraud.
South Korean President Lee ordered a thorough investigation into personal data leaks at financial and public institutions on October 4 local time. At its emergency meeting, the FSC instructed all financial institutions to inspect internet-accessible IT systems, verify authentication and access controls, and share threat intelligence rapidly.
The irony of defensive AI turning against its users
The case carries another implication. South Korean financial institutions have been required by network separation rules to isolate their business systems physically from outside networks. Since June, however, regulators have gradually relaxed those rules to allow AI security testing, expanding eligible companies from 49 to 75 at the start of October.
Rules were relaxed to strengthen defenses through AI, while attackers used AI to breach internal systems. Commentators have suggested that the relaxation may now be reconsidered.
An institute official said open-source AI attack tools such as ARTEX are developed and distributed worldwide, and there is currently no way to restrict them all. As autonomous AI-driven cyberattacks (in Japanese) become real, this case highlights the asymmetry between attackers' low costs and defenders' vast responsibilities. An attacker needs a GitHub account and an LLM API key, while a defender must protect every internal system employees use throughout the organization.
Originally published in Japanese on Joho no Todai.
