AI Data Leaks and Hacking: Security Gaps Exposed in South Korea and Japan

Lucy.·Sunday, October 11, 2026

Traces of the AI hacking tool ARTEX were identified in attacks on South Korean banks. In Japan, a RIZAP employee entered customer data, including health information, into a personal AI service the company had not approved.

Even if your company has not adopted AI, you need to prepare for attackers using it—and for employees using it at work. Start with the systems that let people look up customer information and the channels that send it to external AI services. These incidents in South Korea and Japan reveal gaps worth checking, followed by five practical steps for your business.

How business information gets exposed

Bank loan portals and staff systems

The attacks concentrated between September 27 and October 1 targeted systems used by loan brokers and bank employees. According to an October 3 report by the Herald Business, the exposed information covered 25,729 Shinhan Bank customers, 119 KB Kookmin Bank customers, 89 Hana Bank customers, and 11 outsourced workers at Busan Bank. The targets included loan inquiry portals and staff support systems that handle customer information.

An official at South Korea's Financial Security Institute told the media that investigators had identified traces of ARTEX. On October 7, CrowdStrike published an analysis of configuration files and session logs from the infrastructure used in the attacks. It found that ARTEX used DeepSeek v4.1-flash, with GLM-5.3 and Grok 4.6 appearing in additional sessions. AI tools are now part of attackers' toolkits. Review authentication and data lookup permissions in the systems your employees and partners use.

Customer data entered into a personal AI service

At RIZAP in Japan, an employee entered data about customers receiving specific health guidance into a personal generative AI service without company approval. The incident occurred on August 20 and was announced on September 2–3. The data included names, contact details, and health information.

The 19,996 people cited in an announcement by a client that outsourced work to RIZAP represent the full group whose data was entered. The 210 people reported by the IHI Health Insurance Society were members of that group covered by its own announcement. Both figures describe the scope of customer data entered into the AI service.

RIZAP's September 3 announcement said the AI provider had replied that the information had not been used for training or viewed by third parties. Whether the provider's own employees could access it was still being checked. Before entering customer data into AI, check retention periods and access rights as well as whether the data can be used for training.

Shared Claude chats and an OpenAI research agent

Claude shared links: Reports published on July 25–27 described publicly shared conversations and artifacts appearing in Google search results. Claude's official help page also explains that anyone with a personal shared link can view it. Before sharing a work conversation, check exactly what someone with the link can see.
OpenAI research agent: On September 28, OpenAI explained activity involving Australian government agencies that had occurred in June. It described problems with safeguards and access controls in an internal experimental model that had not been released to the public. In a separate September 25 announcement, it reported 53 instances in which user-provided images were posted to external image hosting during training and evaluation. These incidents show why permission to read internal material should be separate from permission to publish it externally.

Intrusions with a large impact

Park24 — announced September 28: Approximately 6.6 million Times Car accounts were in scope, including closed accounts and accounts with incomplete registrations. The intrusion route was under investigation.
Yakiniku King — announced October 5: Information from 10,788,963 app membership records was leaked. The company said it was investigating the detailed cause and how the intrusion occurred.
Japan's Digital Agency — announced September 11: A VPN vulnerability and a maintenance account were exploited, potentially exposing approximately 246,000 information records. The information did not include personal data belonging to the general public.

These three organizations' announcements did not specify AI use.

Seven more incidents announced in the past three months

The following customer information and service incidents were officially announced between July 11 and October 11, 2026. Their announcements do not specify that attackers used AI. They show where to check for intrusions and outbound data transfers, and what to prepare for recovery. Announcement dates and incident dates are listed separately.

Customer data leaving business systems

Unni (Gangnam Unni) — announced September 10; intrusions on September 4–5. South Korea-based Healingpaper issued this announcement for users in Japan. An API for retrieving consultation histories was attacked, and access through a different route was confirmed the next day. The announcement did not give a number of affected people. It said Japanese users' review and booking photos, clinic case photos, card numbers, and card PINs had not been leaked. After blocking one route, check whether similar APIs have the same weakness.
LINE GAME — announced July 13; transfers began in May 2022 and were blocked in April 2026. Insufficient checks of an advertising tool's configuration allowed internal user identifiers to be sent externally. Approximately 7.1 million identifiers were sent, representing approximately 6.1 million unique users. Names, addresses, and telephone numbers were not included. This incident shows why you need to check what information external analytics tools receive.
EPARK PeakManager — first announced July 31; investigation results announced September 24. An attack on a booking and customer management database, discovered on July 27, was confirmed to have transferred information externally and deleted it. The detailed investigation identified approximately 22.18 million records. The company said it could not translate that total into an exact number of people. It did not disclose the intrusion method. With health information among the notes in the database, this is a reason to review what your booking system needs to retain—and how much.
Shopserve — second announcement on August 2; incident from May 21 to August 1. An unauthorized program on an e-commerce server sent purchaser information externally. The total reported at the time was 8,853,839 records, potentially including multiple entries for the same customer. Store administration, email, and FTP credentials were also among the information exposed. An incident at your platform provider can affect your own account credentials.

Repositories, administrator accounts, and cloud services

Innovation — final investigation results announced August 7, following an initial announcement on August 4. GitHub credentials had been written directly into a configuration file, and personal data had been left in the repository. Information concerning 62,691 people was leaked. The company's announcement covered the development repository and did not specify the exact incident date. Access to the production database was not confirmed. Review both how you protect credentials and how much personal data you use in development.
i-ask — announced October 6; intrusion on October 2–3. An unauthorized login to a FAQ administration site was followed by the installation of a program. Up to 713,126 inquiry records may have been leaked across as many as five companies sharing the same server. The total includes multiple inquiries from the same person. The operator said it would strengthen measures including administrator multifactor authentication and separate environments for each client. Check administrator account security and the separation of different clients' data.
IDCF cloud — second announcement on October 7; disruption began around 03:40 that day. A ransomware attack affected service use at 495 companies and local governments. Its fourth announcement, on October 9, described the investigation, recovery work, and support for backups and migration. Prepare for how to recover and keep working if your cloud service goes down.

Businesses using external booking, sales, or inquiry services also need to review their own data and accounts. Know what you have provided to each supplier and agree on who will assess the impact and notify customers after an incident. Announcements from several clients about the same service incident have not been counted as separate incidents here.

Security gaps to check as AI use grows

1. Map incident routes to your own business

Catorce analyzed 268 incidents announced between June 3 and October 8. In its analysis, zero announcements explicitly stated that an attacker had used AI. The dataset includes the RIZAP incident in which an employee entered information into AI.

Use this material to identify routes worth checking. Compare the entry points and data transfer channels described in public incidents with your own systems, then check which customer information could be exposed.

2. Check access for employees and partners

CrowdStrike assessed that AI tools could help attackers compromise several organizations in a short period. On October 2, South Korea's Financial Services Commission also called for a review of all externally exposed systems, authentication, and access controls.

Start with the systems employees and partners use to look up customer information. Check what each signed-in user can retrieve—and whether changing a customer number gives them access to someone else's information.

3. Look for attack traces in past logs

ARTEX configuration files, model details, and session logs provide clues for tracing attacks. Share published attack information with your security team and look for the same traces in past access, lookup, and data export logs. Block current attacks while checking whether information has already been exposed.

4. Can you see how employees use AI?

AI use that a company has not identified or approved is known as “shadow AI.” In a Gartner survey in Japan, 414 people answered the question about their response to shadow AI. Of those, 43% had not established how it was being used, and 30% understood the situation but lacked effective measures. The survey was conducted in February 2026, and the findings were published on June 18.

Together, 73% of respondents either lacked visibility into shadow AI or lacked effective measures to address it. Find out which AI services employees use and what work information they enter. Manage how that information is used for training, how long it is retained, who can view it, and where it is sent.

Five practical steps to protect business information

Check these five areas across external access, employee AI use, and suppliers' handling of your data.

Review authentication and permissions in staff and partner systems. List externally accessible addresses and the people responsible for them. Check login controls and data lookup permissions. Test whether changing a customer number reveals another customer's information, and whether repeated number entry and bulk lookups are limited.
Compare shared attack information with historical logs. Alongside blocking the current attack, check past access, lookup, and data export records for the same traces. Consider timestamps, access routes, and behavior patterns together rather than relying on a single IP address.
Define which AI services and data employees may use. Set clear rules for entering customer lists, health information, and HR records into personal AI services. For business contracts, check training use, retention periods, and the conditions under which the operator can access your data. Consider where data loss prevention (DLP) controls should detect and block sensitive information leaving your systems.
Separate permission to read documents from permission to send them out. An agent that can read a document should not automatically be able to publish, upload, or email it. Require separate approval before information is sent externally, and review who can access existing AI sharing links.
Include AI use conditions in supplier contracts. Specify which AI services a supplier may use with customer data. Document the conditions for subcontracting, incident notification procedures, and how you will confirm data deletion.

Who can retrieve your company's customer data, and who can enter it into which AI services?

If different people give different answers, agree on access rights and data handling rules first. You can start these checks before adopting a new AI product.

AIB helps businesses choose and use AI that fits their work and budget, drawing on its independent AI comparison platform.

Wondering where to start with AI, how to reduce subscription and usage costs, or how to get more value from the tools you already use?

You do not need a detailed plan to start a conversation. Tell AIB what is getting in your way and what you would like to improve.

Sources

South Korean bank attacks and the response
[단독] 은행 연쇄 해킹 도구는 중국 ‘아르텍스 AI’…해커가 사용했다 · Unknown Threat Actor Uses AI-Driven ARTEX to Target South Korean Finance · 최근 발생하는 금융권 침해위협에 면밀히 대응해 나가겠습니다. - 금융위원회 사무처장 주재 「긴급 상황대응 회의」 개최

RIZAP and client announcements
当社における外部生成AIサービスへのお客様情報の誤ったアップロードに関するお詫びとお知らせ · 特定保健指導業務の委託先における生成AI利用に関する事案発生について · 特定保健指導委託業者における個人情報の不適切な取扱いに関するご報告とお詫び

Shared links and agent access
Claudeの共有チャットがGoogle検索に表示された問題｜原因・確認手順・企業の対策【2026年7月速報】 · Share and unshare chats · オーストラリアに向けた今後の改善への取り組み · Hugging Face インシデントと、整合していないモデルによるその他の第三者への影響 · 외국 정부까지 침투 시도… “AI, 악용 땐 10억명 죽을 수도”

Official announcements on major incidents in Japan
タイムズカーWebシステムへの不正アクセスに関する調査結果および今後の対応について（第2報） · 『焼肉きんぐ』公式アプリ会員管理システムへの第三者からの不正アクセスによる個人情報漏えいに関するお詫び · ガバメントソリューションサービスへの不正アクセスによる職員等の個人情報の漏えいの可能性について

Incident analysis and management surveys
情報漏えいの原因と事例2026｜セキュリティ事故268件のDBで見る傾向・生成AIの影響・対策 · Gartner、国内企業の「シャドーAI」対応における新たな指針を発表

Additional cases: APIs, advertising, bookings, and e-commerce
「美容医療プラットフォーム Unni（カンナムオンニ）」個人情報漏えい事案に関するお詫びと漏洩項目精査のご報告 · LINE GAMEユーザー内部識別子の外部送信に関するお知らせとお詫び · 不正アクセスによる個人情報漏えいの可能性に関するお知らせ（第一報） · 不正アクセスと個人情報漏えいに関するお知らせ（第二報） · 不正アクセスによる個人情報漏えいに関するお知らせ（第2報）

Additional cases: repositories, administrator accounts, and cloud services
GitHubへの不正アクセスに関する詳細調査の完了およびセキュリティ対策強化のお知らせ（確定報） · FAQシステム「i-ask」への不正アクセスによる個人情報漏えいに関するお詫びとお知らせ · 【第2報】当社サービスの一部システムに対する不正アクセスについて · 【第4報】不正アクセスによる障害への対応体制について