VULNpocalypse
HighBreach / campaignActive

ARTEX AI pentest framework abused in data theft from South Korean banks

A suspected Chinese-speaking, financially motivated actor used the ARTEX agentic pentest tool to steal data from South Korean financial firms.

Published
Updated
Exploitation
Yes, CrowdStrike and Zscaler ThreatLabz report data exfiltration from South Korean financial firms, with PII of about 67,000 people per Zscaler

What happened

CrowdStrike reported that a suspected China-based threat actor used ARTEX, a recently released open-source agentic pentesting tool developed in China, together with Anthropic’s Claude model, in a campaign against South Korean financial organizations. The campaign ran from late September to early October 2026 and successfully exfiltrated data. Infosecurity Magazine

Zscaler ThreatLabz describes ARTEX as a framework that uses large language models to automate reconnaissance and exploitation. It says the campaign affected multiple South Korean financial firms and led to the exfiltration of PII belonging to approximately 67,000 individuals. Zscaler ThreatLabz

Key details from the sources:

  • Targets: ARTEX agents probed peripheral internet-facing web applications outside core banking infrastructure, including customer-facing inquiry portals and employee-facing mobile application backends. Zscaler ThreatLabz
  • Techniques: LLM-driven agents exploited vulnerabilities including SQL injection and authentication bypass. Zscaler ThreatLabz
  • Stolen data: full names, phone numbers, annual income figures and loan limits. Zscaler ThreatLabz
  • Named victims: Shinhan Bank and Yegaram Savings Bank reported breaches affecting 25,000 and 40,000 people, according to The Straits Times as cited by Infosecurity Magazine. At one bank the actor reportedly breached a loan progress inquiry service used by financial brokers, and at another an employee mobile work-support system. Infosecurity Magazine
  • AI backends: ARTEX used DeepSeek v4.1-flash as the primary LLM, supplemented by GLM-5.3 from Zhipu AI and Grok 4.6. Infosecurity Magazine
  • Monetization: the actor used Claude Code sessions to seek advice on selling the stolen data and to find Korean-language Telegram channels that trade stolen data. Zscaler ThreatLabz
  • Attribution: CrowdStrike and ThreatLabz both assess with moderate confidence that the actor is a financially motivated Chinese speaker. All attacks were linked to the same IP address, which hosted an ARTEX instance and open directories with Claude Code material. Infosecurity Magazine
  • Guardrail gap: ARTEX’s default blocklist was meant to restrict targeting of government and education domains, but it did not cover financial institutions. Zscaler ThreatLabz
  • Developer response: after public disclosure the ARTEX developer made the project closed source, but earlier copies and forks may remain available. Zscaler ThreatLabz
  • Regulator: South Korea’s Financial Services Commission issued a consumer alert on October 6 warning customers of the hacked companies about phishing and loan scams. Infosecurity Magazine

CrowdStrike said the total number of affected organizations remains unconfirmed. Infosecurity Magazine

CyberScoop reports a different, ongoing ARTEX campaign that CrowdStrike attributes to a China state-backed group it tracks as Vapor Panda. Since October 17 it has targeted more than 200 organizations in at least 14 countries, and about 140 targets were compromised in 11 days. CrowdStrike told CyberScoop that about 70% of targets were confirmed compromised or showed evidence of a likely intrusion. CyberScoop also says it is separate from ARTEX activity that Microsoft attributed to Storm-0940, including a ConnectWise supply-chain attack. CyberScoop

Whether the South Korean bank actor is connected to Vapor Panda is not confirmed in the sources. CrowdStrike assessed the bank actor as financially motivated, while Vapor Panda is described as state-backed.

CrowdStrike’s Adam Meyers said ARTEX could exceed 1,500 commands an hour, against up to 40 per hour for traditional hands-on-keyboard activity. CrowdStrike counted roughly 45 agentic adversary attacks in September and 126 in October. CyberScoop

Am I affected?

  • South Korean financial firms and their partners: you are most at risk if you operate peripheral web apps such as customer inquiry portals, broker-facing services or employee mobile work-support backends. Those were the observed targets.
  • Customers of breached firms: the Financial Services Commission warned them to watch for phishing and loan scams. Infosecurity Magazine
  • Other internet-facing operators: the Vapor Panda activity spans government, defense, telecom, pharmaceutical, semiconductor, nuclear and financial sectors across at least 14 countries. CyberScoop
  • Teams that use open-source pentest harnesses: ThreatLabz advises treating such tools as hostile code. Zscaler ThreatLabz

No CVE, vulnerable product version or patch is named in the sources. The reported weaknesses are application-level flaws (SQL injection, authentication bypass). The Zscaler indicator table is truncated in the text provided, so specific network indicators are not confirmed here.

What to do now

  1. Inventory every internet-facing web app and mobile backend outside your core systems, including partner and broker portals, and test them for SQL injection and authentication bypass. Zscaler ThreatLabz
  2. Enforce server-side authorization on customer records and limit responses to the fields each operation needs. Zscaler ThreatLabz
  3. On partner and employee portals, alert on sequential lookups and unusually high numbers of records accessed per session, and rate limit affected endpoints. Zscaler ThreatLabz
  4. Reduce external exposure, for example by placing applications behind a zero trust access solution, and apply least-privilege user-to-application access to limit lateral movement. Zscaler ThreatLabz
  5. Expect machine-speed scanning, with requests as often as every 10 seconds and multiple attack paths run at once. Make sure you have visibility into endpoints, identities and cross-domain activity. CyberScoop
  6. If you use an open-source pentesting harness, isolate it from production, keep production credentials out of it, enforce scope in configuration and code rather than prompts alone, and review forks before running them. Zscaler ThreatLabz
  7. Retest authorization whenever applications change, and verify users cannot reach records or functions outside their permissions. Zscaler ThreatLabz
  8. If you are a customer of an affected South Korean firm, treat unexpected calls, messages or loan offers with suspicion, as the regulator advised. Infosecurity Magazine

Sources