Summary
- 01What happened
From late September to early October 2026, South Korean financial firms were breached one after another. On 14 October the Financial Services Commission (FSC) put the firms with confirmed breaches at 16 (4 banks, 4 savings banks, 2 capital firms, 2 online investment platforms, 1 e-finance firm, 2 crypto-asset firms and 1 mutual finance body) and the leaked customer records at about 100,000. The fields were names, contact details, resident registration numbers, email addresses and the like; no account passwords or card numbers were confirmed leaked.
- 02What was found
According to CrowdStrike’s analysis, the attacker used ARTEX, an open-source AI pentesting tool developed in China, together with large language models such as DeepSeek, and broke into a loan progress inquiry service used by loan brokers at one bank and an employee mobile work-support system at another. According to the FSC, after the Shinhan Bank incident of 30 September the damage spread to other major firms including KB Kookmin Bank. The attacker has not been identified and is assessed as likely a Chinese speaker with a financial motive.
- 03Why it got through
An externally exposed business inquiry service lets a query through on a matching ID and password alike, whether a person found the credentials or an AI tool brute-forced them.
A pre-execution check means checking a query against the application record before the inquiry service returns data, and keeping a record. Detection means noticing the leak, reporting to the authorities and sharing the attacker's IP addresses. Dates rely on the FSC's releases and CrowdStrike's analysis.
An AI pentesting tool reached several firms
What got through
- What got through
- Authentication to externally exposed business systems, and reads of customer data. According to CrowdStrike, a loan progress inquiry service used by financial brokers was breached at one bank and an employee mobile work-support system at another. At its 2 October meeting the FSC asked every firm to check for paths that reach internal data without authentication, with particular attention to personal-information lookups.
- What reached the accepting side
- In general, an externally exposed inquiry service receives only an ID, a password and the query conditions. It does not receive which broker’s which application the query serves, or whether a person typed it or an AI tool issued it automatically. An AI pentesting tool runs the probing for weaknesses, the login attempts and the data reads automatically, under a person’s instruction.
- Not disclosed
- The intrusion route and vulnerability class at each firm. The FSC’s review of 555 firms lists “missing security patches” and “missing authentication and authorisation and no multi-factor authentication on external systems” as shortcomings, but does not tie them to individual incidents. Whether all 16 firms were hit by the same actor (the FSC says this is under investigation; CrowdStrike says the targets overlap with industry reporting but that the number of affected organisations is unconfirmed). The FSC’s releases do not name ARTEX. Where the leaked data went (CrowdStrike says the actor asked Claude about Korean data-sales Telegram groups, but whether the data was sold is unknown). How much AI accelerated the attacks (Recorded Future says the available evidence does not establish this).
Pre-execution check
Whatever the route, the condition for telling them apart is that, before an externally exposed inquiry service returns customer data, the query is checked as resting on a specific broker’s specific loan application, against an application record kept apart from the ID and password.
Before an inquiry service returns data, check which broker's which loan application the query is for, against the application record rather than the ID and password.
- Who approves
- The customer who applied, and the loan broker handling that application. A query is tied to the application number and issued in a form the broker’s account alone cannot produce, such as the customer’s consent given at application.
- Where it is checked
- The step in which the inquiry service returns data. Even with a matching ID and password, a query not tied to an application record, or beyond the cap on count, period or fields, is held. An employee mobile work-support system is likewise checked against a work request.
- What is recorded
- Which account read whose data under which application. It is kept apart from the inquiry service and used to establish the scope of a leak and to report to the authorities.
Detection worked, and the authorities’ sharing of the attacker’s IP addresses across the sector surfaced breaches at other firms. Even so, about 100,000 records leaked from 16 firms. Cases where an AI agent carried out an intrusion through to exfiltration are SHADOW-AETHER and GTG-1002 (an AI coding agent carried out most of the attack steps). Cases where a vendor’s or provider’s account passed authentication without showing which request it served are Amadeus and Yumirlink. Here, a publicly released open-source tool passed the authentication of an externally exposed inquiry service.
What Lemma does
Lemma builds the pre-execution check into the side where an externally exposed inquiry service returns data. It does not look for attacks by AI tools; it checks, before returning data, that a query rests on a loan application.
Industries and operations to watch
- IndustryFinance
- Banks, savings banks and capital firms that take loan applications and confirm their progress through inquiry services exposed to brokers, agents and partners. The IDs and passwords for those services sit with many outside businesses.
- OperationData access
- Reads of customer data from externally exposed inquiry services.
- OperationAccounts & permissions
- Issuing and revoking broker and employee accounts, multi-factor authentication, and the inventory of externally exposed systems.
Current controls vs Lemma
The route in this case is column 1 (an externally exposed inquiry service’s authentication passed on a found ID and password). Column 2 covers a legitimate broker querying beyond scope with their own account, and column 3 a direct read of the database through a server vulnerability; nothing disclosed says either happened here.
| Control | Authentication with a found ID and password | Out-of-scope queries by a legitimate broker | Direct reads through a server vulnerability |
|---|---|---|---|
| Multi-factor authentication at login | Effective | Not effective | Not effective |
| Security patches and attack-surface reviews | Partial | Not effective | Effective |
| 24-hour monitoring and IP blocking | Partial | Partial | Partial |
| Lemma (pre-execution check) | Effective | Effective | Not effective |
Multi-factor authentication stops logins on an ID and password alone, but the FSC lists “missing authentication and authorisation and no multi-factor authentication” on external systems among the shortcomings found. Patches and reviews close known weaknesses but do not stop a query with valid credentials. Monitoring and blocking worked here once the authorities shared the IPs, but after the leak. Lemma checks, before returning data, not whether the credentials are valid but whether the query rests on an application.
Note: Lemma’s check works only when it is built into the step that returns data from the inquiry service and the application records are managed apart from the brokers’ accounts. If the server is breached and the database read directly, data can be read without passing this check. It also passes if a broker’s account is taken over and an application record is created with it. The table compares what each control checks; it does not show whether this incident would have been stopped. Lemma does not detect or prevent intrusions, nor does it tell attacks by AI tools from others.
Three things to confirm about externally exposed business systems
- Have you listed every inquiry service exposed to brokers, agents and vendors, and is there any that returns customer data on an ID and password alone?
- Do you cap queries by count, period and fields, and hold queries beyond the cap before returning data?
- When the authorities or the industry share an attacker’s IPs or methods, can you check your own records the same day for whether queries were made?
Facts and timeline
- Parties: 16 South Korean financial firms (banks: Shinhan, KB Kookmin, Busan, Hana; savings banks: Yegaram, Welcome, DB, OK; capital firms: Hyundai Capital, KDB Capital; online investment: PFCT, Mouda; e-finance: Brightfix; crypto assets: Streami, Pravving; mutual finance: Korean Federation of Community Credit Cooperatives) (FSC, 14 October). Supervisors: the Financial Services Commission (FSC), the Financial Supervisory Service (FSS) and the Financial Security Institute.
- What happened: According to the FSC, 16 firms were confirmed breached around the time of the Shinhan Bank incident, with different systems compromised and different data leaked at each. About 100,000 customer records (names, contact details, resident registration numbers, email addresses and the like) leaked; no data that could lead directly to fraudulent payments, such as account passwords or card numbers, was confirmed leaked. Firms are still checking against the shared attacker IPs, so the count may rise.
- CrowdStrike’s analysis (7 October): Open directories on the attacker’s infrastructure held ARTEX configuration files and Claude Code session histories and memory files. The ARTEX instance used DeepSeek v4.1-flash as its main large language model, and GLM-5.3 (Zhipu AI) and Grok 4.6 were also used in Claude Code sessions. A Chinese-language pentesting prompt was included. The setup was two servers: a main server on a Hong Kong IP and a separate IP running ARTEX. The actor asked Claude “where threat actors typically sell Korean data breach information” and for help finding Korean data-sales Telegram groups. The activity ran from late September to early October. The actor is unattributed and assessed, with moderate confidence, as likely a Chinese speaker with a financial motive.
- Recorded Future’s analysis (TAG-217): Between 17 and 22 August the actor tested remote-access infrastructure and scanned the internet; by 26 September the environment held AI tools and records tied to reconnaissance, traffic proxying through residential connections, payload delivery and examination of payment processes. The actor appears opportunistic, trying many targets until finding valuable ones. The available evidence does not establish how much AI accelerated the attacks.
- Shinhan Bank (FSC, 2 October): A data leak incident occurred on 30 September, after which cyberattacks on other major firms including KB Kookmin Bank caused further damage. The FSC said it would inspect firms on site after incident reports and oversee consumer protection and compensation. The breached system and the count at each firm have not been disclosed.
- FSC review: 555 firms submitted a 12-item checklist by 8 October; 425 reported “appropriate” and 130 reported shortcomings (concentrated in e-finance and online investment). The main shortcomings were missing security patches, poor management of end-of-support systems, no 24-hour real-time monitoring, and weak attack-surface management including missing authentication and authorisation and no multi-factor authentication on external systems.
Glossary
Pentesting tool — A tool for testing whether one's own systems can be entered the way an attacker would. ARTEX is an open-source one developed in China that hands the probing and intrusion steps to a large language model (an agentic tool).
Loan broker — In South Korea, an outside intermediary (대출모집인) that gathers loan applications on a bank's behalf and checks their approval status through an inquiry service the bank provides.
Timeline
- 2026-08-17 to 22
- Tests of remote-access infrastructure and internet scanning on the attacker’s infrastructure (Recorded Future).
- Late September 2026
- Attacks on South Korean financial firms begin (CrowdStrike).
- 2026-09-30
- Data leak incident at Shinhan Bank (FSC).
- 2026-10-02
- The FSC holds an emergency response meeting chaired by its secretary general and orders inspections of externally exposed systems, checks for authentication and authorisation weaknesses, sharing of attacker IPs and threat information, and a checklist.
- 2026-10-04
- An emergency inspection meeting for the whole financial sector, chaired by the FSC chairman.
- 2026-10-07
- CrowdStrike publishes its analysis that the actor used ARTEX and large language models.
- 2026-10-08
- 555 firms submit their checklist results.
- 2026-10-14
- At its second emergency meeting the FSC discloses 16 breached firms and about 100,000 records, the review results (425 “appropriate”, 130 with shortcomings) and its preventive policy (a full inspection, vulnerability tests at least twice a year, an inventory of external access points with stronger authentication, CEO and CISO accountability, threat-information sharing and AI-based inspection).
The facts rely on the FSC’s releases (2 and 14 October; primary), and the analyses by CrowdStrike and Recorded Future (independent analysis). Shinhan Bank’s own notice could not be confirmed, and the breached system and count at each firm have not been disclosed. The FSC’s releases do not name ARTEX; CrowdStrike says the targets overlap with industry reporting while leaving the number of affected organisations unconfirmed. Whether all 16 firms were hit by the same actor has not been disclosed. The leaked count is “about 100,000”, with no breakdown by firm. What has not been disclosed does not mean there were no controls. This Brief does not evaluate the firms’ responses. Written as of 6 November.
After disclosure
- FSC, FSS and Financial Security Institute: Shared the attacker’s IPs across the sector and reviewed 555 firms with a 12-item checklist. They plan a full inspection of all firms’ security management, vulnerability tests at least twice a year for large firms, an inventory of external access points with stronger authentication and access control, clearer reporting and accountability for CEOs, CISOs and CIOs, a threat-information sharing scheme with AI-based threat-level reporting, AI-based shared monitoring for smaller firms, and more effective sanctions. The FSS and the Institute are investigating whether AI was used in the attacks.
Sources
- Primary, regulator’s releaseFinancial Services Commission「최근 발생하는 금융권 침해위협에 면밀히 대응해 나가겠습니다」
- Primary, regulator’s releaseFinancial Services Commission「금융권 침해사고 재발방지를 위해 금융보안 관리체계를 전면 강화하겠습니다」
- Primary, regulator’s release, EnglishFinancial Services Commission “Authorities Hold Meeting on Cybersecurity Management in Financial Sector to Bolster Anti-Hacking Measures”
2 more (reporting, analysis)
- Independent analysisCrowdStrike “Unknown Threat Actor Uses AI-Driven ARTEX to Target South Korean Finance”
- Independent analysisRecorded Future “TAG-217 Targets South Korean Financial Services with ARTEX”
References: “The last layer left in AI-era cyber defense”