Hackers used AI tools to break into banks' business apps. What a mid-size company should check

What happened

Security firm CrowdStrike published research on Oct. 7 about a campaign against South Korean financial organizations that ran from late September to early October 2026 and ended with data stolen. At one bank, the attacker got into a loan-progress lookup used by outside brokers. At another, an employee mobile work app.

The attacker used ARTEX, an open-source "agentic" penetration-testing tool built in China. Penetration testing is what security teams do to find holes before attackers do. ARTEX automates it with AI. In this case, CrowdStrike says, it ran mainly on a DeepSeek model, with other models used through Anthropic's Claude Code. The attacker also asked Claude where stolen Korean data is usually sold.

CrowdStrike hasn't named the attacker. It assesses, with moderate confidence, that the person is a Chinese speaker after money, based on the tool and the Chinese-language prompts. Its key finding is that AI tooling can let one financially motivated attacker run several break-ins in a short time.

The wider damage was large. Reuters reported that at least nine South Korean banks have disclosed or been reported as targets. Shinhan Bank said personal information of about 25,000 customers was compromised; KB Kookmin Bank said 119. Police opened an investigation. The Korea Times, citing Yonhap, noted that the attacker's identity, the full extent of the breaches and the amount of stolen data remain unconfirmed. CrowdStrike itself says the number of affected organizations is unconfirmed.

One detail is worth knowing: ARTEX's own GitHub page says it is for personal learning and research, and should not be used against live online systems, according to Reuters. That didn't stop anyone. Free, capable tools are now one download away.

AI didn't invent a new attack. It let one person run several at once.

Why a mid-size company should care

It's easy to read "South Korean banks" and move on. Look at what was actually hit: a web portal for outside partners and an app for employees. Almost every 40 to 400 person company has both. A dealer or customer portal. A vendor login. Remote access for the sales team. A phone app for techs in the field.

Big banks have security teams that watch those systems all day. Most mid-size manufacturers, distributors and service companies have one IT person, or an outside provider who checks in when something breaks. What AI changes is the math on the attacker's side. Probing a company's apps for weak spots used to take skill and hours. If a tool does much of it automatically, smaller targets become worth the time.

The AI labs see it coming too. On Oct. 9, Axios reported that executives at OpenAI, Anthropic and other labs are privately rehearsing their response to a major AI incident. Many insiders expect one within six to 12 months, and the most feared is a large cyberattack on banking, internet access or utilities, as reported by Decrypt. OpenAI said such scenarios are not treated as inevitable.

What doesn't change

CrowdStrike's report doesn't describe any new kind of hole. It names no new software flaws and gives no special advice, which tells you something: this was ordinary hacking, done faster. The defenses that worked last year still work. They just can't wait.

The one new item is AI itself. Here the attacker used AI coding tools as a hacking assistant. As more companies connect AI to their own email, portals and ERP, those connections become something to protect too. Treat an AI tool's login like an employee's: its own account, only the access its job needs, and a record of what it did. We wrote more about that in what Microsoft's CEO calls an AI emergency brake.

Where the other side has a point

A fair skeptic would say this story is being stretched. The attribution is "moderate confidence." Reports differ on how many banks were hit and by whom. The breaches happened in Korea, not Texas. And security companies have a business reason to make every threat sound new. All true.

It's also fair to say that the same AI tools help defenders. Security teams use them to find holes first, which is exactly what ARTEX was supposedly written for. The point isn't panic. It's that the slow, cheap advantage small companies used to have, that nobody bothered with them, is shrinking.

Five checks for this month

1. List everything that faces the internet. Customer and dealer portals, vendor logins, remote access, webmail, field apps. If you can't list them, start there.
2. Turn on multifactor sign-in for every one of them. CISA says users who enable it are significantly less likely to get hacked, because a stolen password alone isn't enough. Start with email and remote access.
3. Patch the outside-facing apps first. CISA keeps a public list of flaws attackers are actively using. Ask your IT person or provider to check it against what you run.
4. Make sure someone sees the alerts. Logins from strange places, many failed sign-ins, big downloads. A log nobody reads doesn't help.
5. Give any AI tool its own limited login. If you connect AI to your email, portals or ERP, it should have only the access its job needs, so a mistake or a break-in can't reach everything.