← Blog

Korea Banks: ARTEX AI Tool in ~68K Customer Leaks

Mon Oct 12

Share on X

Between late September and early October 2026, a cluster of South Korean financial institutions began confirming customer and staff data leaks that English and Korean press rolled into a sector-wide headcount of roughly 68,000 people — a sum built from each institution’s own disclosures, not a single regulator census, and one that shifts slightly when outlets consolidate overlapping categories. The largest verified slice is Shinhan Bank, which put loan-related personal data for about 25,729 customers in the open after outsiders reached a loan-agent inquiry path. Smaller but equally confirmed notices followed from KB Kookmin Bank (119 customers via an employee mobile support tool), Hana Bank (89 customers via an ODS sales support system), and additional lenders and savings banks named in the same news cycle. Investigators at the Korea Financial Security Institute and reporting in outlets including the Korea Herald business desk describe traces of ARTEX AI, a Chinese-developed open-source large-language-model penetration-testing framework published on GitHub under the Autumn-27 account, on infrastructure tied to the wave. To be clear about what that means: ARTEX is being discussed as a tool in human-directed intrusions, not as an autonomous AI agent that chose targets and exfiltrated data without an operator behind the keyboard.

This article is a verified-campaign guide to the October 2026 Korean banking leak wave. It links each major confirmed institution to its BreachHistory catalog row, separates attested customer counts from press aggregates, and walks through what ARTEX AI actually is — without turning “China-developed open source” into unsupported state-sponsored attribution. Primary English anchors used here include Yonhap, GovInfoSecurity, and the Herald piece above.

What happened: a wave, not one breach

Headline writers often compress a week of separate intrusions into “the Korean bank hack,” as if one gang downloaded a monolithic 68,000-row CSV from a core banking mainframe. The verified picture is messier and more instructive. Multiple victims disclosed on adjacent days. Entry paths cited in bank statements and regulator-facing coverage point at employee tools, partner-facing inquiry services, and sales-support systems — the plumbing that lets loan solicitors, mobile relationship managers, and ODS desk staff look up customer context quickly. Those systems sit beside, not inside, the hardened retail internet-banking channel customers use on their phones.

Shinhan’s disclosure remains the gravitational center. On October 1, 2026, the bank confirmed that unauthorized external parties reached services through abnormal methods that bypassed authentication, exposing loan-application context for tens of thousands of customers. Yonhap and follow-on Korean reporting placed the active leak window on Wednesday, September 30. CEO Jung Sang-hyuk apologized publicly and pledged full compensation while the Financial Supervisory Service sent an on-site inspection team — a signal that regulators treated Shinhan as the first domino, not the full story.

Within hours, peer institutions tested their own environments and published smaller counts. KB Kookmin Bank confirmed 119 customers affected through an employee mobile support system. Hana Bank confirmed 89 customers after attackers touched its ODS sales support system, stressing that ODS is separate from internet and mobile banking and that transactional ledger data was not described as leaked in the notices BreachHistory indexed. BNK Financial Group disclosures reported in the same press cycle included 11 contract workers at BNK Busan — a tiny number that still matters because outsourced staff often hold operational credentials. Savings institutions also surfaced: reporting commonly cites roughly 40,000 customers at Yegaram Savings Bank and about 2,200 corporate customers at Welcome Savings Bank, alongside mentions of Hyundai Capital in sector roundups. Treat Hyundai Capital and the savings-bank figures as institution-reported or press-consolidated unless you have that firm’s own notice in hand; the three megabank rows below are the ones with the clearest company confirmations in the catalog.

When journalists summed institution reports into “about 68,000 people,” they were doing arithmetic on public disclosures, not quoting a unified FSS spreadsheet. If one outlet folds Shinhan’s ~25,729 together with Yegaram’s ~40,000 and the KB/Hana/BNK slices, you land near 68,000; if an editor deduplicates categories or waits for a revised Shinhan assessment, the banner number moves. BreachHistory records each victim separately; this campaign post is the map.

Confirmed victims and catalog links

Use these canonical paths when you share or bookmark a single incident from the wave:

The pattern across those three confirmations is more important than any one headcount. Attackers did not need to defeat consumer-grade app attestation on millions of phones simultaneously. They walked through doors built for employees and partners — inquiry APIs, mobile support consoles, ODS lookup desks — that were never marketed to customers but still stored identifiable loan and KYC context.

Timeline: late September intrusion, early October disclosures

  1. September 30, 2026 (Wednesday) — Shinhan’s attested leak date for loan customer data; emergency blocking of external IPs and suspension of affected services follow in bank statements reported by Yonhap and Aju Press.
  2. October 1, 2026 (Thursday) — Shinhan CEO apology, FSS on-site inspection, customer lookup tools announced; Yonhap English coverage frames the incident as hacking of borrowing details and contact information.
  3. October 1–2, 2026 — KB Kookmin and Hana publish confirmations; Financial Services Commission and FSS convene emergency sector meetings reported in Korea Herald and Korea Times roundups.
  4. Same window — Additional savings-bank and capital-finance names appear in consolidated press pieces; BNK’s outsourced-worker leak surfaces as part of the cluster.
  5. Early October investigations — Korea Financial Security Institute work and Herald reporting describe ARTEX AI artifacts on examined systems; GovInfoSecurity summarizes international attention on “AI-assisted” wording while noting human operation.

The public clock is fast by U.S. breach-notification standards. Korean lenders apologized and stood up lookup hotlines within days of the attested intrusion window, which helps customers but also means forensic details — exact exploit chains, dwell time, full field inventory — may arrive in later FSS bulletins. If you are waiting for a perfect final census before changing passwords or refusing a suspicious loan call, you are optimizing for the wrong variable.

What data was exposed — institution by institution

Shinhan Bank loan-agent system

Shinhan’s confirmed set centers on loan application and solicitation context: customer names, phone numbers, annual income, and estimated loan eligibility or borrowing limits. Korean afternoon reporting added narrower but high-sensitivity slices — on the order of dozens of resident registration numbers and connecting-information identifiers — rather than claiming every one of the ~25,729 rows included full national IDs. Income plus a plausible limit is social-engineering fuel. A caller who already knows your name, phone, and approximate borrowing ceiling can impersonate a Shinhan loan officer without stumbling over basic facts.

KB Kookmin employee mobile support

KB’s 119-customer notice describes exposure through an internal employee mobile support system — the kind of tool relationship staff use away from a branch desktop. English-language reporting tied to the wave lists names, phone numbers, addresses, and encrypted or protected resident registration data among the fields in scope for affected customers. The count is small, but the architectural lesson matches Shinhan: support mobility expands the attack surface.

Hana Bank ODS sales support

Hana confirmed resident registration numbers, names, addresses, emails, phones, and employer names for 89 customers reached via ODS sales support. The bank’s public emphasis — ODS is not internet banking — is a useful reminder that “online banking was safe” can still coexist with meaningful PII leaks elsewhere in the group.

Other institutions in press aggregates

Roundups that drive the ~68,000 figure also name Yegaram Savings Bank (~40,000 customers in press arithmetic), Welcome Savings Bank (~2,200 corporate customers), BNK Busan (11 contract workers), and Hyundai Capital. BreachHistory indexes those where separate verified rows exist; if you only read one English article, verify which entities issued their own notices versus which were mentioned in a journalist’s sum. Do not treat Hyundai Capital’s mention in sector press as equivalent to Shinhan’s CEO-signed apology unless Hyundai publishes matching confirmation.

What was not exposed — based on current attestations

Across the megabank notices BreachHistory used, institutions drew a bright line between partner/employee support systems and core consumer internet or mobile banking channels. Hana explicitly stated that financial transaction data was not leaked in its ODS incident. Shinhan’s October 1 reporting focused on loan-inquiry fields rather than claiming full deposit account numbers or card PANs for the entire ~25,729 population. Absence of a claim in day-one notices is not a forever guarantee — FSS inspections can widen scope — but it is the correct baseline for customer comms until a regulator or bank says otherwise.

Nothing in the ARTEX AI reporting indexed here asserts that attackers drained SWIFT terminals or rewrote ledger balances. The verified harm model is identity and loan-context theft suitable for follow-on fraud and phishing, not a cinematic “empty every account” headline.

How the attacks worked — partner paths and authentication bypass language

Public descriptions from Shinhan and sector coverage converge on architecture, not a single named CVE. Outsiders reached services that should have required strong authentication and instead encountered abnormal methods that bypassed authentication — phrasing that fits broken access control on a partner inquiry API, stolen or shared solicitor credentials, or a support console exposed to the open internet without matching the bank’s consumer-app hardening.

KB’s path label points at employee mobile support, which implies either compromised staff devices, stolen support accounts, or a vulnerable mobile gateway that trusted internal sessions too broadly. Hana’s ODS label points at sales-support workflows where staff legitimately query customer records during product conversations — a workflow that stores RRNs because Korean KYC practice demands it.

Reporting on the wave repeatedly notes that attackers did not need to mass-hack customer phones running retail banking apps. That distinction matters for threat modeling inside other banks: the next incident may again be the “small” system your board deck never lists beside the mobile app logo.

ARTEX AI: what investigators are actually describing

ARTEX AI entered English-language coverage as a provocative phrase — “AI hacked Korean banks” — but the underlying artifact described by the Korea Financial Security Institute and Herald business reporting is a Chinese open-source LLM-assisted penetration testing project associated with GitHub user Autumn-27. GovInfoSecurity’s summary of the South Korean situation likewise frames the tool as something that may have helped attackers explore systems, not a self-directed agent that independently executed the entire campaign.

In practical terms, ARTEX-class tooling is closer to an automation assistant for a human pentester or intruder: suggesting next steps, scripting repetitive reconnaissance, or organizing exploit attempts against services exposed on a network. The LLM does not magically bypass authentication by itself; an operator still chooses targets, interprets results, and exfiltrates data. Korean investigators examining servers tied to the wave reported Chinese-language strings and tooling consistent with that project on examined infrastructure. That is evidence about tooling presence, not a completed attribution docket naming a sponsor or a motive.

What this is not, and BreachHistory will not claim without primary attribution sources:

  • Not proof that “AI decided” to attack Shinhan autonomously.
  • Not proof of Chinese government involvement merely because the open-source project is China-developed — open-source security tools circulate globally, and presence of a repo origin country is not a geopolitical verdict.
  • Not a bank-confirmed root-cause line replacing authentication-bypass findings; megabank notices emphasize access paths and exposed fields, while ARTEX discussion lives in institute and press investigation layers as of early October.

Security leaders should still take the story seriously. LLM-assisted recon lowers the skill floor for finding forgotten partner APIs. It does not replace the need for MFA, network segmentation, and logging on employee support systems — it may accelerate an attacker who already has a foothold or stolen credentials.

Who is at risk

Shinhan loan applicants and recent borrowers face the widest verified population. Anyone who interacted with loan solicitation in the affected window should assume name and phone exposure at minimum, and should watch for callers who reference income or limits.

KB and Hana customers in the confirmed small sets face higher identity integrity risk where RRNs and addresses were attested. Employer names in Hana’s ODS set enable “your HR department and your bank” combo scripts.

Savings-bank and corporate customers named in ~68,000 rollups should follow each institution’s official lookup channels rather than this article’s summary alone.

Outsourced BNK workers whose records leaked face spear-phishing aimed at people who can reset vendor access — a classic supply-chain pattern even when the count is only eleven.

Employees at Korean financial firms broadly should expect mandatory password rotations, VPN policy changes, and phishing drills referencing “FSS emergency inspection” themes. Those emails will be duplicated by criminals within days.

Industry and campaign context

South Korea’s financial sector runs hot on digital lending and partner-sourced loan volume. Inquiry services for solicitors exist because branch staff cannot manually re-key every field when a fintech partner submits an application. That business logic creates databases that are not “core banking” in marketing terms but are still rich in regulated personal data. The October 2026 wave is the latest proof that regulators and CEOs must inventory those databases with the same rigor as mobile banking gateways.

Internationally, the episode landed beside other 2026 finance breaches where the soft door was a vendor or employee console — not the consumer app icon. Readers comparing incidents should cross-read BreachHistory rows for Shinhan, KB Kookmin, and Hana rather than treating “68,000” as a single breach ID.

Woori Bank and NH NongHyup were reported in the same Korea Times and Herald cycle as institutions that saw attack attempts without confirmed customer data loss in the first wave of English summaries. That nuance matters for executives planning peer comparisons: not every Korean logo in a headline graphic suffered a confirmed census.

What companies and regulators said

Shinhan led with a CEO apology, compensation pledge, customer lookup tooling on its website, and planned Super SOL app features for self-check — paired with FSS on-site inspection. KB and Hana followed with smaller confirmed counts, IP blocking, emergency teams, and public apologies stressing customer anxiety. Sector regulators convened emergency meetings reported in Korean business press, signaling coordinated scrutiny rather than treating each notice as an isolated PR problem.

Investigation-layer statements attributed to the Korea Financial Security Institute — relayed through Herald and international trade press — introduced ARTEX AI as a forensic finding on examined systems. Those statements should be quoted as investigator reporting, not as text lifted from a Shinhan customer PDF, unless and until a bank notice repeats the tool name.

Phishing and fraud models to expect after this wave

Expect three lanes of secondary crime:

  • Loan impersonation — SMS or Kakao messages referencing Shinhan or peer banks, citing real income or limit details, pushing “re-verify to release funds” or “pay a processing fee.”
  • RRN and employer phishing — calls to Hana-scale victims who hear their employer name and a partial address, then get asked to “confirm” full resident registration numbers for a compliance audit.
  • Fake FSS or “digital arrest” scams — preying on panic after national headlines about AI and Chinese tooling; legitimate regulators do not demand remote-desktop sessions or cryptocurrency.

Hang up on unsolicited loan officers. Call back using the number on the back of your card or an official bank app entry — not a number in the text message. Never install remote-support software because a stranger says your account is “frozen due to the ARTEX hack.”

Action items if you may be in scope

  1. Use official bank lookup tools first. Shinhan stood up website menus and app features for exposure checks; peer banks published their own channels in the October notices. Ignore QR codes from social media threads.
  2. Freeze or monitor credit where RRNs were attested. For Hana’s 89 and any notice listing resident registration numbers, treat identity theft as a concrete risk, not a theoretical one.
  3. Rotate credentials on any loan broker portal you used in September 2026, especially if you reused passwords across fintech partners.
  4. Enable phishing-resistant MFA on email and brokerage accounts tied to your phone number — leaked phones become account-recovery vectors.
  5. Document losses if fraud occurs. Shinhan’s compensation pledge creates a paper trail; keep police reports and bank ticket numbers.
  6. Warn household members. Family members may receive convincing calls referencing your loan application even if they bank elsewhere.
  7. Employees: report odd support tickets. If your bank employer asks you to run penetration scripts or install unknown GitHub tools on production-adjacent hosts, escalate through official security channels — legitimate FSS inspections do not arrive as zip files in Telegram.

Checklist for security teams at other banks

  1. Inventory every partner inquiry API, ODS console, and employee mobile support gateway that can return customer PII without traversing retail MFA.
  2. Require step-up authentication and IP allowlists on those systems even if internal users complain about friction.
  3. Log and alert on bulk query patterns from solicitor or support accounts — ARTEX-style automation may show up as faster-than-human browsing.
  4. Segment support networks from general office IT; assume compromised laptops will probe internal URLs.
  5. Pre-draft customer notices and regulator call trees so the next peer breach does not force improvised wording about AI.

Open questions investigators still owe the public

Early October reporting leaves deliberate gaps. Exact exploit chains for Shinhan’s authentication bypass are under FSS review. Full overlap between ~25,729 and any later revised Shinhan total may change. Whether ARTEX artifacts appear on every victim or only on a subset of examined hosts may clarify as the Korea Financial Security Institute publishes more. Credential stuffing and purchased access broker involvement were discussed in trade press as possibilities — not confirmed replacements for the partner-system narrative. Hyundai Capital’s role, if any, awaits institution-specific notices beyond name mentions in roundups.

Journalists and bloggers should label each sentence with its layer: company notice, regulator inspection, or institute forensic reporting. Mixing layers is how “68,000 AI-hacked accounts” memes outrun the verified Shinhan/KB/Hana troika.

Canonical records and authoritative sources

BreachHistory campaign summary (this page will publish at slug korea-banks-artex-ai-wave-october-2026 when ingested). Per-incident canonical URLs:

Outbound references:

If a regulator or bank updates counts after FSS inspections conclude, the catalog rows above are the authoritative place BreachHistory will amend — not reposted screenshot threads.

Bottom line for readers searching “Korean bank data breach 2026”

Yes, the wave is real, verified at multiple named institutions, and large when press sums institution reports into roughly 68,000 people. The harm is primarily personal and loan-context data exfiltrated through employee and partner systems, not a confirmed meltdown of every consumer mobile banking app in the country. ARTEX AI is part of the investigation story as an open-source LLM pentest tool allegedly found on related infrastructure — wielded by humans, not a rogue autonomous agent, and not by itself proof of nation-state sponsorship. Start with your bank’s official lookup, assume loan-themed phishing for months, and use the three megabank canonical links if you need shareable, incident-specific records.

Blogs you may like