Why AI-Assisted Attacks Go Through The Side Door

The attackers who broke into South Korea’s largest banks in late September did not come in through internet banking.

They went in through a service for loan agents, an app for employees and a sales-support system. The data did not leave through the core transaction networks that draw the security attention.

That is the useful part of this story, and it has little to do with artificial intelligence. AI-assisted attacks make it cheap to try every door at once. Whether you lose data then depends on which doors you forgot you had, and AI-assisted attacks are very good at finding those.

Where the AI-assisted attacks actually landed

The facts are still settling, but the pattern is clear across the reporting. Shinhan Bank lost data on about 25,700 customers through a service used by loan agents. KB Kookmin Bank lost 119 records from an internal mobile work-support system for employees. Hana Bank lost 89 records from an employee sales-support system. At BNK Busan Bank, the exposed data belonged to 11 outsourced developers rather than customers.

The Korea JoongAng Daily reports that the attacks then spread beyond the large banks to savings banks and consumer finance companies. Across the cases, the targets were systems facing the internet but built for staff, agents and contractors. Customer-facing internet and mobile banking were not where the breaches happened.

The systems nobody ranks first

Every organisation has these systems. Think of the reseller portal, the status page for brokers, the field app for the sales team and the login for the agency that maintains the website. A business unit built each one to solve a real problem, and each one is reachable from the internet.

The Korea Herald describes the targets as relatively less-secure systems, and that is the honest description. They sit outside the customer channels that attract audits, penetration tests and board attention. They also tend to hold more than their purpose requires, because nobody asked them to hold less.

The BNK Busan case adds a further detail. The people exposed were contractors. Your attack surface includes the data you hold about the people you work with, not only the people you sell to.

How much of this was AI

Less settled than the headlines suggest. Security researchers and investigators reported traces of ARTEX AI, an open-source penetration-testing system built on a large language model and published on GitHub. However, the reporting differs on how far investigators have confirmed its use. BleepingComputer noted on 5 October that authorities had not yet confirmed it in the Shinhan breach, while the Herald Business reported that the Korea Financial Security Institute had found evidence the tool was used.

The authorities’ working reading is more instructive than the attribution. As the Kyunghyang Shinmun reports, they see the attacker trying multiple financial firms and breaching the ones with weaker defences, rather than targeting particular companies. That is what automation changes. It does not need to be clever. It only needs to be thorough, and the tooling for thoroughness is now free.

So the question for an organisation is not whether AI-assisted attacks will reach it. It is whether your weakest internet-facing system is one you would recognise if someone named it.

What the regulator asked for

The Financial Services Commission did not respond with an AI policy. At an emergency meeting on 2 October, it ordered banks and card companies to identify every IT asset and service exposed to external access and to examine its vulnerabilities and access controls. The checks had to cover every externally accessible system, whether or not it serves customers.

Firms were also told to minimise the information those systems expose and to look for any route into internal information that bypasses authentication. In other words, the regulator asked for an inventory, a data diet and an authentication check. None of those require any knowledge of AI.

That response is worth borrowing. It meets AI-assisted attacks at the level where an organisation actually has control, which is its own estate.

Applying the lesson before AI-assisted attacks arrive

You do not need to be a bank to run the same exercise. You need answers to four questions about every system that someone outside your network can reach:

• Existence. Is it on a list, including the portals built for partners, agents, contractors and staff?
• Ownership. Which named person owns it today, and who decides when to switch it off?
• Data. What does it hold or reach, and does it need all of it?
• Authentication. Is there any path from it into internal information that skips a proper login?

In practice the exercise stalls at the first question, because AI-assisted attacks scan everything while inventories list only what someone remembered. Customer channels appear on the architecture diagram. The broker portal built three years ago by the commercial team often appears on none. If your inventory relies on what the IT department commissioned, it will miss what the rest of the business bought.

Ownership is the board question

Security teams can scan an estate. They cannot decide that the loan agents’ portal should hold less data, or that the business should retire the reseller login. Those are business decisions, and they need an owner who can make them.

That is why defending against AI-assisted attacks belongs in front of the board rather than only in the security plan. AI-assisted attacks reward organisations that know what they run and punish those that have let the edges grow unattended. The Korean banks are now producing that inventory under regulatory pressure, in the middle of an incident. The cheaper moment to produce it is before one, rather than after AI-assisted attacks have drawn up the list for you.

So, for the next leadership meeting: could anyone in the room name every system your partners and staff can log into from outside, and the person who owns each one?