CrowdStrike links ARTEX and Claude-related infrastructure to suspected South Korean financial-sector attacks

South Korean financial organizations faced a series of cyber incidents in late September and early October 2026, including unauthorized access to Shinhan Bank services and the acquisition of customer information. Separately, CrowdStrike said it found ARTEX configuration files, Claude Code session histories and Claude memory files on infrastructure associated with a campaign targeting South Korean financial organizations.

The findings create a link between attacker-controlled infrastructure and AI-related tooling, but they do not publicly establish that ARTEX or Claude directly conducted the intrusions. The identity of the operator also remains unconfirmed.

What was confirmed

Shinhan Bank disclosed that an external party accessed certain services and obtained customer information. Its public filing did not quantify the number of affected customers and did not identify the attack method.

South Korea’s Financial Services Commission held an emergency response meeting on October 2, 2026. The commission ordered financial firms to inspect externally exposed systems, strengthen authentication and access controls, and share threat intelligence. The FSC release documents a sector-wide response; it does not itself name Shinhan Bank.

Yonhap and other Korean financial-sector reporting described the incident as affecting approximately 25,000 Shinhan customers and involving suspected credential stuffing. Those details should be treated as media-reported claims rather than facts established in Shinhan’s public filing.

What CrowdStrike found

CrowdStrike said it associated the campaign with infrastructure that included the Hong Kong-based host 38.244.50[.]120, an ARTEX instance and an exposed Claude Code document. Researchers also found Claude Code session histories and Claude memory files on infrastructure linked to the targeting of South Korean financial organizations.

The ARTEX instance used DeepSeek v4.1-flash. CrowdStrike said GLM-5.3 and Grok 4.6 appeared in additional Claude Code sessions. The company assessed the actor as likely Chinese-speaking and financially motivated, with moderate confidence. That assessment is a threat-intelligence judgment, not a confirmed identification of a person or organization.

CrowdStrike’s findings concern infrastructure and artifacts associated with the campaign. They are not, by themselves, a complete victim-side forensic chain proving that the exposed tools caused each reported breach.

What ARTEX is designed to do

ARTEX is an open-source, multi-agent autonomous penetration-testing system. Its stated functions include automating reconnaissance, vulnerability discovery, attack-path planning, tool execution and verification.

That design makes the presence of ARTEX configuration files relevant to understanding how an operator may have organized activity. However, the verified evidence does not establish which actions the system performed, whether it reached a particular bank’s systems, or whether it was responsible for the unauthorized access disclosed by Shinhan.

The same distinction applies to the Claude-related material. The evidence refers to attacker-run Claude Code sessions or related artifacts found on associated infrastructure. It does not show that Anthropic’s Claude service was compromised, nor does it prove that Claude performed the intrusions.

Why the wording matters

Describing the event as “Claude agents and ARTEX were used in attacks on South Korean banks” would overstate what is publicly established. The available evidence supports a narrower conclusion: CrowdStrike linked ARTEX and Claude-related files and infrastructure to suspected activity targeting South Korean financial organizations, while South Korean authorities and Shinhan separately confirmed financial-sector incidents.

Those are related strands of evidence, but they are not the same as a public forensic finding that the tools caused the breaches. The distinction is important because vendor threat-intelligence reporting may identify infrastructure, files and observed tradecraft without proving every step of a victim’s intrusion.

It also avoids conflating Claude Code session artifacts with Anthropic-operated agents. The verified material describes sessions and files found in an environment associated with the suspected campaign, not an intrusion into Claude itself.

Response and remaining uncertainty

The FSC’s emergency measures focused on sector-wide defenses: inspection of externally exposed systems, stronger authentication and access controls, and threat-intelligence sharing. These instructions followed the disclosed incidents and were directed at financial firms broadly.

Several key questions remain unanswered in the public evidence. Shinhan’s filing does not state how many customers were affected or how the unauthorized access occurred. The approximately 25,000 figure and suspected credential-stuffing description come from Korean media reporting, not that filing. CrowdStrike’s assessment also does not publicly identify the operator, and the verified material does not support naming an individual.

The evidence therefore supports reporting both developments together—with clear attribution—without treating the association between AI tooling and attack infrastructure as proof that those tools directly caused the bank incidents.