Korean Bank Attacks | AI-Written Résumé Leads Investigators to Suspect
機械翻訳 / Machine-translated
A
AI Friends
2日前・19分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Would it surprise you to learn that the clue to identifying the perpetrator was a "résumé" the attacker had an AI write for them? Here is a breakdown of a case in which an attacker's use of AI left a trail of digital footprints during attacks on South Korean banks.
From late September 2026, multiple South Korean financial institutions experienced a series of data breaches. According to ITmedia, Reuters reported that at least nine banks were targeted.
The scale of damage varied by institution. Based on reporting by the Korea JoongAng Daily, the figures are as follows:
Woori Bank and NH NongHyup Bank were also targeted, but no leaks of customer information have been reported. Across all reports, it has been stated that financial transaction data was not accessed.
South Korea's Financial Services Commission convened an emergency meeting and instructed financial institutions to inspect all externally accessible IT assets. According to BleepingComputer, local media also reported that President Lee ordered a thorough investigation on October 4.
The turning point came with an investigation by U.S. cybersecurity firm CrowdStrike. According to an analysis published on October 7, an open directory — a folder whose contents anyone could view — was found on a server associated with the attack's IP addresses.
Inside, the following items were left behind:
The attacker used two servers, with a Hong Kong-based server at the center. ARTEX was run on a separate server, and nine proxies (relay servers used to conceal identity) were also employed.
The attacker had left their work logs in a state where anyone could view them.
ARTEX is an AI-powered penetration testing automation tool published on GitHub in 2026 by a Chinese security engineer. Penetration testing refers to the process by which defenders proactively search for weaknesses in their own systems.
ARTEX connects to an external LLM (an AI capable of producing human-like text) and delegates the task of finding vulnerabilities to the AI. According to ITmedia, the tool's public page restricts its use to personal learning and code research, and requests that it not be used for penetration testing on live systems.
The AI at the center of this attack was not Claude. According to CrowdStrike, the AI operating behind ARTEX was DeepSeek's "DeepSeek v4.1-flash" from China's DeepSeek. As supplementary tools, China's Zhipu AI's "GLM-5.3" and U.S. xAI's "Grok 4.6" were also used.
The attacks were driven by the combination of ARTEX and DeepSeek. Claude Code was a tool the attacker used for research and document creation.
The Claude Code logs reveal the attacker's activities. According to CrowdStrike, the attacker made queries such as the following:
This suggests the attacker had plans to sell the stolen information. From this, CrowdStrike believes financial motivation is likely.
Another request provided material that shed light on the suspect's identity. The attacker attempted to have Claude write a "security researcher's résumé" incorporating the results of the current attacks as professional achievements.
The request reportedly contained personal information about the individual. According to ITmedia, the attacker introduced themselves as "a 26-year-old residing in Maoming, Guangdong Province, China," and even included a Telegram username and educational background.
However, CrowdStrike has stated that while this information is "likely to be the individual's own," it cannot be confirmed at this time. There are also reports that the age and date of birth listed in the logs are inconsistent. When Reuters contacted the phone number found in the report, the man who answered said he was unaware of the matter.
The "26-year-old" in the headline is nothing more than a self-reported claim written by the person themselves. It is safer not to treat it as an established fact.
In this case, AI expanded the scope of what a single attacker could accomplish. Adam Meyers, who oversees adversary tracking and analysis at CrowdStrike, told ITmedia: "With the power of AI, a single person can target large numbers within a very short period of time."
Here is a comparison of conventional attacks versus AI-assisted attacks like this one:
| Category | Conventional Attacks | AI-Assisted Attacks (This Case) |
|---|---|---|
| Vulnerability discovery | Manually examined one system at a time by a person | ARTEX delegates the search to AI |
| Number of targets | Limited by available manpower | Can target large numbers in a short time |
| AI used | None | Multiple, including DeepSeek, GLM, and Grok |
The activity period in this case was approximately one to two weeks from late September to early October. Within that short window, damage spread across multiple banks.
Note that the full picture of the damage has yet to be established. CrowdStrike itself has stated that the total number of affected organizations has not been confirmed. Please also be aware that figures vary across different reports.
This is not someone else's problem for companies in Japan either. Imagine an IT administrator at a small or medium-sized enterprise. Suppose they have several business systems accessible from outside the company that have been left unattended.
An attacker using AI can rapidly probe such entry points one after another in a short period of time. Unlike with manual attacks, there is no guarantee they will give up partway through.
In this incident, it was reported that at Shinhan Bank, identity verification for a service used by loan intermediaries was bypassed, and at Hana Bank, an internal business support system for employees served as the point of entry. Even at large organizations, peripheral systems became the entry points.
According to a headline from a related ITmedia article, Japan's National Cybersecurity Secretariat has also issued a warning. The Personal Information Protection Commission has cited measures such as strong multi-factor authentication (a system requiring verification beyond just a password) and EDR (a system for detecting suspicious activity on computers) as urgent countermeasures. Please refer to the official announcements from each agency for details.
The three things to review at your own organization first are:
According to BleepingComputer, South Korea's Financial Services Commission has similarly instructed financial institutions to inspect externally accessible IT assets and review their authentication measures.
No. According to CrowdStrike's analysis, the AI at the core of ARTEX's operations was DeepSeek. The Claude Code logs were found as evidence that the attacker used it for research and résumé writing. The article does not touch on what Claude's responses were.
That remains unknown. The self-introduction was found in the request for the résumé, but CrowdStrike is avoiding any definitive conclusions. Their assessment is that the individual is "likely a Chinese speaker with financial motivation," stated with moderate confidence.
The tool's public page restricts its use to personal learning and code research, and requests that it not be used for penetration testing on live systems. The way it was used here violated those terms.
The basics are: do not reuse passwords across banks and membership services, and enable multi-factor authentication. If you receive a suspicious message, contact the official support channel listed on the organization's website rather than any number provided in the message itself.
Start by creating a list of your organization's "systems accessible from outside the company" and checking whether any of them have weak identity verification.
This article is a cross-post from AI Friends.
