CYOPS · Cyber Operations

Korean bank intrusions went through the side doors, and an agentic pentest tool walked them

Intrusions at several South Korean lenders hit broker and staff-facing services, not core banking. CrowdStrike links the activity to the open-source ARTEX agent, but victim counts, technique and attribution remain unsettled.

The intrusions that hit South Korean banks between late September and early October did not breach core banking. They went through side doors: a loan-status lookup built for outside loan brokers, an employee mobile work-support app, a sales-support system [2][3][4]. The tool linked to them is an open-source agentic penetration-testing framework, ARTEX, and the operator left session logs on exposed servers [1]. The lesson for defenders is less about AI and more about tier-two web services that nobody treated as part of the perimeter.

What happened

Shinhan Bank confirmed a breach on 30 September and disclosed it on 1 October. KB Kookmin, Hana and BNK followed on 2 October [3]. The Financial Services Commission (FSC) held emergency meetings on 2 and 4 October and ordered firms to inspect every externally accessible system, including services not aimed at customers [3][4][5]. Firms were told to finish checks and close gaps by 8 October [5]. Police opened a formal investigation, and President Lee Jae Myung ordered a thorough inquiry [5][6].

The entry points were peripheral. At Shinhan, the attacker bypassed identity verification in a mobile service that loan agents use to track applications [3]. At KB Kookmin the route was an employee mobile work-support system; at Hana it was the ODS sales-support system [2][3]. BNK Busan cited insufficient session validation on its web pages [5]. The Financial Security Institute (FSI) said all of the attacks hit employee- or partner-facing systems, and that the attacker queried data rather than taking control of systems [2]. Hana and KB said no transaction data was affected [3]. Regulators had confirmed no customer financial losses as of 6 October [5].

Exposed fields at Shinhan included names, phone numbers, annual income and calculated loan limits, plus 66 resident registration numbers and 97 identity-verification identifiers [3]. That combination is useful for loan-themed voice phishing, which is why the FSC issued a consumer alert [5][7].

Where the figures disagree

The numbers are still moving, and the sources do not agree.

  • Institutions: the Kyunghyang Shinmun counted nine affected firms, including two unnamed online lenders [4]. American Banker reported at least seven [5]. The Cloud Security Alliance (CSA) notes outlets have given five, seven or nine [12]. The FSI said far more institutions were attacked than suffered incidents [2].
  • People affected: the FSI’s confirmed total on 3 October was just under 26,000 [2]. RuntimeWire, citing officials, reported data tied to about 68,000 people across seven firms [9]. The gap is consistent with later disclosures, including about 40,000 people at Yegaram Savings Bank as reported by Infosecurity Magazine [7], but no single official tally reconciles them.
  • Shinhan alone is reported as about 25,000 [3], 25,727 [4] and 25,729 [2] records.
  • KB Kookmin is reported as 119 customers [3][10] and, separately, as 99 customers plus 20 current and former employees [5].
  • Technique: Yonhap described credential stuffing, while Dong-A Ilbo described enumeration of random customer numbers over about 30 hours from 28 September; American Banker notes the conflict and that the phone-check bypass is unexplained [5]. The Kyunghyang Shinmun also describes random-value probing [4].

The ARTEX link and its limits

The FSI told the Herald Business on 3 October that its log analysis pointed to ARTEX, and that a human directed it [2]. Yonhap reported a server used in the attacks carried an HTML page title associated with ARTEX [6]. CrowdStrike then published infrastructure analysis on 7 October. It found open directories on a Hong Kong-hosted server that exposed Claude Code session histories, Claude memory files and ARTEX configuration files, with a second server hosting the ARTEX instance [1].

According to CrowdStrike, ARTEX ran on DeepSeek v4.1-flash as its primary model, with GLM-5.3 and Grok 4.6 used in other sessions [1]. The operator also used Claude to ask where Korean breach data is sold and to find Korean Telegram data-sale groups [1][8]. CrowdStrike assessed with moderate confidence that the operator is a Chinese speaker and financially motivated, and did not attribute the activity to a named adversary [1]. It said personal details found in one session likely belong to the actor but could not be definitively linked [1]. This desk does not repeat those details; press reports naming an individual are unconfirmed, and the account holder has denied involvement [8][13].

Caution is warranted. Banks and regulators had not publicly confirmed ARTEX use as of 5 and 6 October [5][6]. A tool trace on a server does not prove the tool performed each intrusion, and AhnLab’s reported count of about 600 internet-hosted ARTEX instances shows how common the software is, not how many attackers exist [9][12]. Attack IPs came from at least eight countries, including Korea, the US, Japan and Hong Kong [4]; another report put routing at more than 20 IPs in more than 10 countries [11]. The FSI noted that IPs alone cannot identify the attacker [4], and Korean commentators cautioned that the tool’s language does not by itself indicate the operator’s nationality [10].

The tool’s status is also contested. The developer said on 8 October that ARTEX would no longer be updated and would move to closed source, citing misuse [8][13][14]. CSA reported the project was removed from GitHub [12]. When this desk checked on 13 October, a repository page at the original address was publicly reachable with a README limiting use to authorised testing and research [15]. Either way, copies already downloaded remain in circulation [14].

What it means

This was not a sophisticated intrusion. CrowdStrike’s own assessment is that AI tooling let a financially motivated operator run multiple intrusions within a short time span [1]. The FSI’s assessment is that customer-facing channels had been hardened heavily while internal and partner systems were managed less rigorously [2]. An agentic scanner does not need a novel exploit when a broker lookup service trusts a phone check and exposes credit data to people who are not bank staff [5].

The market reaction was material. Demand deposits at the five largest banks fell from 675.5 trillion won on 30 September to 655.4 trillion won by 11 October, although the Korea Times did not attribute the whole fall to the breaches [14]. Banks are now announcing larger security budgets and AI-based red-team programmes [14]. Regulators are also weighing the incidents against a relaxation of network-separation rules for AI security testing, which had expanded from 49 to 75 firms [2].

What to do

  • Inventory every internet-reachable service, including broker, agent, contractor and staff portals. Treat them as perimeter, not back office [2][5][12].
  • Re-test authentication on those services against enumeration and credential stuffing. Enforce server-side session validation on every request [5].
  • Minimise the personal credit data that intermediaries and staff can view, in line with the FSC’s 4 October order [5].
  • Rate-limit and alert on high-volume, systematic querying of lookup endpoints from few sources [12].
  • Ingest CrowdStrike’s published indicators and search historical logs, but expect the operator to rotate infrastructure [1][2].
  • Prepare customer communications for follow-on phishing that uses leaked income and loan-limit data [5][7].

The cyber operations reading

Agentic pentest frameworks are now commodity infrastructure, and takedowns of the source repository do not withdraw capability already deployed [12][14]. The operator here was exposed by poor operational security, not by detection at the banks [1][13]. Asian financial regulators, including those in Malaysia and the wider ASEAN region, should read the Korean case as a test of their own partner-facing services. The relevant control is not an AI detector. It is knowing which low-tier systems exist, who can reach them, and what data they return.

Sources

Every R3KONX article cites its primary material. 15 sources, in order of first citation. Links open the original publication.

  1. Unknown Threat Actor Uses AI-Driven ARTEX to Target South Korean Finance CrowdStrike · 2026-10-07
  2. Exclusive: Chinese AI tool ARTEX used in wave of bank hacks, probe finds The Herald Business · 2026-10-03
  3. Hana, BNK report data breaches as hacking hits Korean banks The Korea Herald · 2026-10-02
  4. The Korean financial sector was breached by a two-month-old Chinese AI The Kyunghyang Shinmun · 2026-10-04
  5. AI-linked hacks hit Korean banks through loan-agent sites American Banker · 2026-10-06
  6. South Korea probes bank breaches amid suspected AI-powered attacks BleepingComputer · 2026-10-05
  7. Chinese Hacker Deployed AI in Campaign Against South Korean Banks Infosecurity Magazine · 2026-10-08
  8. ARTEX AI Pentesting Tool Used in Data Theft Attacks on South Korean Financial Firms The Hacker News · 2026-10-08
  9. AhnLab finds ARTEX on 600 IPs as police probe South Korean bank breaches RuntimeWire · 2026-10-06
  10. [EXPLAINER] How AI emerged as new threat in Korea's bank hacking crisis The Korea Times · 2026-10-07
  11. ARTEX Named in AI Hacking of Korean Banks Seoul Economic Daily · 2026-10-08
  12. ARTEX: Open-Source Agentic Pentesting Tool Linked to South Korean Bank Intrusions Cloud Security Alliance · 2026-10-09
  13. South Korean Banks Hit by an Autonomous AI Pentesting Agent, and CrowdStrike Found It Because the Operator Left the Sessions Open Forkast · 2026-10-11
  14. Banks race to ramp up AI defenses as deposits fall after cyberattacks The Korea Times · 2026-10-11
  15. Autumn-27/ARTEX repository GitHub · 2026-10-13

Researched and written by the R3KONX analysis desk from the cited primary material. Korean regulator notices could not be retrieved directly, so regulatory actions are cited from Korean and international press reporting of them; victim counts are provisional and differ between sources. Corrections to event@r3konx.asia.

R3KONX 2027

Work in this domain? So does the programme

4–6 May 2027, World Trade Centre Kuala Lumpur. Curated talks, hands-on training and The BattleGrid.