Hackers targeting South Korean financial institutions used artificial intelligence systems developed by several technology companies to conduct cyberattacks and steal data, according to findings from cybersecurity firm CrowdStrike.The investigation identified the use of Anthropic’s Claude Code, DeepSeek, xAI’s Grok and Zhipu AI alongside ARTEX, a Chinese-developed automated penetration-testing framework. The attacks occurred between late September and early October, affecting financial organisations already facing investigations into compromised customer information.
CrowdStrike identified infrastructure associated with the campaign after discovering publicly accessible directories containing AI session histories, configuration files and operational records. These materials allowed investigators to reconstruct aspects of the attacker’s activities and establish connections between different intrusions.
The company assessed with moderate confidence that the unidentified operator was Chinese-speaking and financially motivated. Its assessment relied on Chinese-language instructions, the selection of hacking tools and evidence recovered from systems controlled by the attacker.
The findings have intensified scrutiny of artificial intelligence in financial-sector cybersecurity, particularly the ability of individual operators to combine automated testing frameworks with commercially available language models.
CrowdStrike established that ARTEX used DeepSeek v4.1-flash as its principal language-model backend. Additional Claude Code sessions involved Zhipu AI’s GLM-5.3 and Grok 4.6, demonstrating how multiple AI services could support different activities within one operation.
Investigators identified two servers central to the campaign. Infrastructure based in Hong Kong functioned as the operator’s principal system, while another server hosted the ARTEX framework believed to have been used against South Korean financial institutions.
The exposed directories contained Claude Code session records, ARTEX configuration documents and memory files. These records revealed instructions for automated penetration testing and helped investigators establish overlaps between the identified targets and financial institutions reporting breaches.
The compromised systems reportedly included a loan inquiry service used by financial brokers and a mobile work-support application used by bank employees. Such systems provide operational access without necessarily forming part of a bank’s principal transaction-processing infrastructure.
Shinhan Bank, KB Kookmin Bank and Hana Bank were among the institutions affected by the broader series of cybersecurity incidents. Authorities have been examining the intrusions and assessing the extent of information exposed.
However, the precise number of institutions compromised through ARTEX remains unconfirmed. CrowdStrike has also not established that every breach under investigation involved the same operator or identical methods.
South Korean authorities initiated investigations following disclosures of customer information and concerns about weaknesses in financial institutions’ security arrangements. President Lee Jae Myung called for stronger protection of critical systems as investigators examined possible AI involvement.
The campaign has also prompted questions about the responsibilities of developers whose legitimate security-testing software is subsequently misused.
ARTEX was designed to automate penetration-testing activities, including identifying vulnerabilities and coordinating security assessments. Its integration with language models enabled users to direct technical tasks through natural-language instructions rather than relying entirely on manually executed commands.
Following disclosures linking the framework to the attacks, its developer withdrew the publicly available project and announced that further updates and support would cease. Previously distributed copies and derivative versions nevertheless remained available.
Security specialists distinguish the misuse of such frameworks from evidence that the underlying AI companies knowingly supported criminal operations. The available findings establish the attacker’s use of their technologies, not participation by the developers in the intrusions.
CrowdStrike also discovered evidence suggesting that the operator considered selling stolen information. Claude Code session records showed requests concerning marketplaces for compromised Korean data and Telegram groups associated with trading such material.
Those requests indicate possible commercial intentions, although investigators have not established whether the stolen information was successfully sold or whether the attacker received payments.
Another session contained a request to prepare a security researcher’s curriculum vitae, incorporating claims about penetration-testing activities. Personal information entered into the conversation suggested a possible connection to Guangdong province in China.
Investigators cautioned that inconsistencies in the supplied information prevented definitive identification of the operator. No named hacking organisation has been conclusively linked to the campaign, and the evidence does not establish government sponsorship.
CrowdStrike identified nine proxy addresses associated with the ARTEX-related activity, alongside the two principal servers. The company documented these technical indicators to assist security teams investigating potentially connected intrusions.
The records also suggested that the operator accessed DeepSeek through an intermediary API service, rather than necessarily connecting directly to the model provider.
Topics
Technology