CrowdStrike: AI Hacking Tool Let One Attacker Breach Korean Banks

A suspected Chinese-speaking attacker stole more than 25,000 customer records from Shinhan Bank alone, according to CrowdStrike — and the security firm says AI-powered tooling made the multi-bank campaign possible for what was likely a single operator.

The intrusion hit multiple South Korean financial institutions, not just one. CrowdStrike, which disclosed its findings in a report covered by The Decoder, identified the attacker through linguistic and operational indicators as a likely Chinese speaker. The company did not name every affected bank in the disclosed material, but it confirmed Shinhan Bank as a major victim, with the tally of stolen customer records there exceeding 25,000.

The weapon at the center of the case is ARTEX, an open-source tool that automates penetration testing using AI models including DeepSeek and GLM-5.3. ARTEX is not a bespoke implant built by a state laboratory. It is openly available, and it wraps large language models into a workflow that can probe, enumerate and exploit targets at machine speed. That combination — public availability plus automated offensive capability — is what CrowdStrike flags as the story's core problem.

Why does one attacker with ARTEX change the threat model?

CrowdStrike's assessment is blunt: the case shows how AI tools can let a single person pull off massive breaches. That sentence carries weight because it inverts a long-standing assumption in enterprise defense.

For years, security teams sized their adversaries by headcount. Opportunistic criminals worked alone and hit soft targets. Coordinated campaigns against hardened financial institutions implied a crew — reconnaissance specialists, exploit developers, operators, money mules. Staffing was a natural ceiling on how much damage one actor could do.

ARTEX removes part of that ceiling. The tasks that once required a team — mapping a target's infrastructure, triaging vulnerabilities, chaining exploits into a working intrusion path — are exactly the tasks penetration-testing AI models are built to automate. One person supervising an automated pipeline can now generate the volume of activity that previously signaled an organized group.

For South Korean banks, the stakes are direct. Financial institutions hold exactly the kind of concentrated customer data — identities, account relationships, contact records — that turns a single breach into tens of thousands of individual exposures. The 25,000-plus records taken from Shinhan Bank represent the downstream cost of one operator running one open-source tool.

What is ARTEX, and why does 'open source' matter here?

ARTEX is described by CrowdStrike as an open-source automated penetration-testing tool that uses AI models such as DeepSeek and GLM-5.3. Each element of that description matters.

Open source. The code is freely available. There is no procurement barrier, no underground purchase, no developer relationship to trace. Any actor who can download a repository can arm themselves.

Automated penetration testing. Penetration testing itself is a legitimate, defensive-adjacent discipline — security professionals use it to find holes before criminals do. ARTEX's significance is that it points that same automated capability at live targets and runs it without a professional in the loop requiring time, salary or expertise per step.

AI-native. The tool leans on large language models, including DeepSeek and GLM-5.3, to drive its attack logic. These are general-purpose models repurposed for offensive work. The attacker did not need to train anything; the model capabilities arrived ready to use.

The result is an asymmetry. Defensive teams still cost what they always cost. Offensive capacity has become downloadable.

What did the attacker actually take?

At Shinhan Bank, the confirmed figure is more than 25,000 customer records. CrowdStrike reported the theft as part of its investigation into the broader campaign against South Korean financial institutions.

The attribution signal — a suspected Chinese-speaking attacker — comes from CrowdStrike's analysis of the operator behind the intrusions. The company framed the identity assessment as a suspicion grounded in its investigative indicators, not as a confirmed government attribution, and the disclosed reporting does not establish state sponsorship.

What does this mean for banks and regulators?

The incident lands at a moment when policymakers are still arguing about how to govern offensive AI capabilities. CrowdStrike's case study offers the concrete data point that debate has lacked: not a lab demonstration, but a real campaign against real banks, with a real record count.

Three implications follow directly from the disclosed facts.

• Volume without headcount. A likely single attacker reached multi-institution scale. Defensive planning that assumes small actors stay small no longer holds.

• Commodity tooling, institutional targets. An open-source tool with public AI models was sufficient to engage hardened financial-sector victims — the attack surface is not limited to unprepared small businesses.

• Record-level impact. More than 25,000 customer records from one bank alone quantify the blast radius of a single automated intrusion.

For security teams, CrowdStrike's findings argue for monitoring that keys on behavior rather than on assumed adversary profiles. If one person can produce the operational footprint of a team, then sizing the threat by its apparent scale will mislead.

What comes next?

CrowdStrike presents the case as evidence of a shift already underway, not a hypothetical: AI tooling has lowered the barrier to large-scale breaches far enough that a lone, suspected Chinese-speaking attacker could hit multiple South Korean financial institutions and walk away with over 25,000 customer records from Shinhan Bank. As automated penetration tools like ARTEX and the models behind them — DeepSeek, GLM-5.3 — remain publicly available, the question for defenders is no longer whether AI-assisted intrusions will scale, but how quickly incident-response capacity can adapt to adversaries who no longer need a team.