The Side Door Problem

Every building has a front door, and every security budget knows where it is. It gets the best lock, the camera, the guard and the annual inspection. In application security, that's the customer banking app, the checkout, the crown jewels. It's what the pentest covers.

Every building also has side doors. The loading dock. The employee entrance. The window the contractors use. Nobody forgets they exist. They just never make the inspection list, because inspections cost money and a side door doesn't feel like the risk.

For a long time that was a reasonable bet. Side doors were expensive for attackers too.

What happened in Korea

From late September into early October, one operator breached at least seven Korean financial firms. CrowdStrike's analysis points to a single, financially motivated actor working for a few weeks, likely using an open source AI pentest agent called ARTEX alongside commercial AI coding agents.

The interesting part isn't the tool. It's which doors it tried.

Core payment systems were reportedly untouched. The doors that gave way were a loan inquiry service used by brokers, an employee mobile work app and a sales support system. Each faced the internet, each handled customer data, and each sat well below the crown jewels on anybody's list.

Side doors, every one.

Kay was a threat analyst for Korea's financial sector institutes before joining Xint, and frames it simply: the places attackers aim haven't changed in ten years. What changed is the cost of finding them and getting in.

The price of a door

Think about what an attacker's budget used to look like. Working a broker portal meant a skilled person spending days on reconnaissance, credentials and the application's logic. That cost forced a choice, so most attackers made the same choice defenders did and went for the obvious door.

An agent doesn't have to choose. It enumerates what's exposed, replays logins leaked from other breaches, maps the APIs behind them and checks whether one customer's session can read another customer's data. When it gets blocked, it tries another door. The cost per door is now compute and patience.

The defender's cost per door didn't move. It's still a pentest, priced by the app and the day. So the inspection list, once a reasonable compromise between two expensive sides, is now a map of the doors only one side can afford to try.

The American version

None of this is a Korean problem. US lenders run broker and dealer portals. Insurers run agent portals. Every company of any size runs employee apps, partner APIs, regional sites that outlived their migrations, and that one marketing microsite with a login.

Compliance pays for the front door. PCI and NYDFS both expect an annual pentest, and in practice both are satisfied by a tight scope around the systems that matter most. That's rational. It's also exactly the list of doors an attacker with an agent tries first.

It's worth being precise about what's new. Open source AI attack tools have been around for a while; Villager showed up on PyPI in 2025 and was downloaded about ten thousand times. What's new is evidence: one operator, several banks, a few weeks, documented. The economics everyone predicted showed up in a bank's logs.

What defenders have that attackers don't

Start with the basics. Most of these doors opened with reused passwords, so MFA and rate limiting on every external login closes a lot of them. But a password only opens the door. How much walks out depends on what's behind it: at one bank, about 25,700 people's records came out of a loan inquiry service, one lookup at a time.

The good news is that the same economics are available to defenders, and defenders hold three advantages.

• The keys. You can try every door you own, including the ones behind a login, with permission.

• The blueprints. An attacker has to infer your business logic from the outside. You can read the source.

• The floor plan. You know, or can find out, where every door is. The attacker is still walking the perimeter.

What you need is testing cheap enough that every door makes the inspection list, and precise enough that a hundred doors don't turn into a hundred triage queues. That second part matters more than people expect. Checking a hundred side doors and getting back ten thousand weaknesses isn't progress. You need vulnerabilities: findings proven exploitable, with the steps to reproduce them.

That's what we built Xint to do. It attacks the running application the way an outsider would, reads the source where you can share it, and reports what it can prove. In Korea, one bank we work with rechecked every affiliate within a day of the first breach going public.

Change the question

Most security programs ask whether the front door held. After Korea, the better question is how many side doors we have, and when anyone last tried each one, us or them.

If the answer is never, you already know who tries it first.

We spent years hardening the front door because that's where the budget pointed. AI just made every side door worth trying. The inspection list should change with it.