Did an AI pentest tool called ARTEX breach South Korean banks, and what should I check?
Published Updated Editorial policy
CrowdStrike says one actor used ARTEX, an open-source AI pentest tool, plus AI models against South Korean financial firms in late September 2026. Banks and authorities had not confirmed the tool in early reports. What is clear is that ordinary side systems were reachable, so inventory yours.
Who should act: Anyone who runs partner portals, inquiry services, employee mobile back ends or sales tools that face the internet, banks first.
Update log
- : Published.
What happened
The Kyunghyang Shinmun reported on 4 October that between 27 and 30 September attackers leaked data from Shinhan Bank, KB Kookmin Bank, Hana Bank, BNK Busan Bank, Yegaram Savings Bank and Hyundai Capital. It also named Welcome Savings Bank and two P2P lenders. Shinhan lost 25,727 items of personal data, and Kookmin and Hana lost data on 119 and 89 customers. The paper says the attackers hit weaker systems: a loan-progress inquiry service for brokers at Shinhan and an employee mobile work-support system at Kookmin. It presumes an automated brute-force approach, with random values tried to find valid customer numbers.
South Korea's Financial Services Commission met on 4 October. It told financial firms to inspect every externally accessible IT system, customer-facing or not, cut unnecessary exposure, check authentication and access controls, share threat information and submit inspection results (BleepingComputer, 5 October).
CrowdStrike (7 October) found open directories on attacker servers holding Claude Code session histories, ARTEX configuration files and Claude memory files. ARTEX ran DeepSeek v4.1-flash as its main model, with GLM-5.3 and Grok 4.6 added to the Claude Code sessions. CrowdStrike names no adversary and rates a Chinese-speaking, financially motivated actor as moderate confidence. It says the number of affected organizations is unconfirmed.
The counts differ. Kyunghyang names nine. The Next Web (8 October) says police cover seven and Reuters counts at least nine banks. The CSA note (9 October) cites five, seven and nine in different outlets. The note adds that ARTEX traces do not prove the tool did the intrusions, because credential stuffing could explain some.
Who is affected
- Customers of the named lenders. Exposed fields include names, contact details, annual income and loan limits, so expect loan-themed phishing. The commission issued a consumer alert on 6 October (Infosecurity Magazine).
- Any organization with internet-facing side systems: broker portals, inquiry pages, staff mobile back ends, sales support and microsites.
- Teams that assume the core system's login protects everything. A security firm's chief executive quoted by Kyunghyang said the main systems had more than passwords and the peripheral ones did not.
What to do now
- List every public hostname and service you run or a vendor runs for you, not only customer-facing ones. Start from DNS and certificate records.
amass enum -passive -d example.com - Give each one an owner and a decision: remove it, put it behind a VPN or an allowlist, or harden it.
- Test your own lookup endpoints for enumeration. Sequential or guessable IDs that answer differently when valid are what the reported attack walked. Run this only against systems you own.
seq 1000 1050 | xargs -I{} curl -s -o /dev/null -w '%{http_code}\n' https://staging.example.com/status/{} | sort | uniq -c - Add per-source and per-identifier rate limits and bot detection, and alert on one source making thousands of lookups in minutes.
- Search logs since 27 September for the ten addresses CrowdStrike published. Nine are proxy addresses, so investigate matches before you block them.
grep -c -F -f ioc-ips.txt /var/log/nginx/access.log - Remove or encrypt national ID numbers, income and loan limits on ancillary systems, so a breach there yields less.
What should have been in place
- A current inventory of internet-facing systems, each with a named owner.
- The same authentication strength on side systems as on the core platform.
- Rate limits and detection tuned for machine-speed probing.
- Data minimization on ancillary systems, so they hold less to steal.
- Authorized tests with agentic tools, in a segregated environment under written rules, and only from vetted builds. The CSA note recommends this.
Questions people ask
How many banks were breached in the ARTEX campaign?
Unconfirmed. CrowdStrike says the count is unconfirmed. Kyunghyang names nine firms, The Next Web says seven in the police case and nine banks per Reuters, and the CSA note cites five, seven and nine. Treat any single number as provisional.
Was Claude Code used in the attacks?
CrowdStrike found Claude Code session histories and memory files on the attacker servers. It says ARTEX itself used DeepSeek v4.1-flash, with GLM-5.3 and Grok 4.6 supplementing the Claude Code sessions. The CSA note says the sources do not say which model did the scanning.
Can I still download ARTEX?
The developer made the project closed source and stopped updates after real-world use, BleepingComputer reported on 10 October, and English and Korean derivatives of the old code exist. Use only vetted builds, for authorized tests.
Is it proven that ARTEX did the intrusions?
Not by the victims. CrowdStrike's link rests on the operator's own session logs. Banks and authorities had not confirmed ARTEX use when BleepingComputer reported on 5 October.
Digiton's production review lists every externally reachable system and how each one authenticates, and the checks it runs are on the vibe coded app audit page.
How this brief was made: researched from the sources below, drafted with AI assistance and checked against each source on the dates shown. See the editorial policy.
Sources
- CrowdStrike, unknown actor uses AI-driven ARTEX against South Korean finance, 7 October 2026, read 26 October 2026. https://www.crowdstrike.com/en-us/blog/unknown-threat-actor-uses-artex-to-target-south-korean-finance/
- Kyunghyang Shinmun, the Korean financial sector was breached by a two-month-old Chinese AI tool, 4 October 2026, read 26 October 2026. https://www.khan.co.kr/en/article/202610042320007
- BleepingComputer, South Korea probes bank breaches, 5 October 2026, read 26 October 2026. https://www.bleepingcomputer.com/news/security/south-korea-probes-bank-breaches-amid-suspected-ai-powered-attacks/
- BleepingComputer, ARTEX AI and Claude agents used against South Korean banks, 10 October 2026, read 26 October 2026. https://www.bleepingcomputer.com/news/security/hacker-used-artex-ai-and-claude-agents-to-target-south-korean-banks/
- Cloud Security Alliance, research note on ARTEX and the bank intrusions (AI-assisted, not formally reviewed), 9 October 2026, read 26 October 2026. https://labs.cloudsecurityalliance.org/wp-content/uploads/2026/10/CSA_research_note_artex_agentic_pentest_bank_breaches_20261009-csa-styled.pdf
- The Next Web, CrowdStrike links the South Korea bank hack to AI tools, 8 October 2026, read 26 October 2026. https://thenextweb.com/news/crowdstrike-south-korea-bank-hack-suspect-china
- Infosecurity Magazine, Chinese hacker deployed AI against South Korean banks, 8 October 2026, read 26 October 2026. https://www.infosecurity-magazine.com/news/chinese-hacker-ai-korean-banks/
More security briefs
How these briefs are sourced and corrected. All security briefs
Building something that has to hold up in production?
Start with a free 25-minute scoping call. If it fits, the first paid step is a production review: five working days on your real repository, scored against ten layers, with a costed fix list.
Book a 25-minute scoping call