Before your company adopts AI, ask how far one lost login reaches

The attacker behind the attacks on South Korean banks could not lock their own server. With AI lowering the bar for attacks, what a company can design is its blast radius: how much one compromised account can reach.

The attacker CrowdStrike tracked could not lock their own server. The instruction file they gave their AI (CLAUDE.md) and their conversations with it sat on the open internet, readable by anyone without a login.1 Someone that careless with the basics still used AI tools to go after major South Korean banks, in a wave of attacks that CrowdStrike says ended in stolen data.

01What happened

The Korean bank breaches and ARTEX: what is confirmed and what is not

Fact

Between late September and early October 2026, several South Korean banks and financial companies found their systems had been breached. Media counts put it at seven or more institutions, with the names, phone numbers, annual incomes and loan limits of about 68,000 people exposed.23

What was breached was not the banks’ core ledgers but the systems around them: one bank’s service where loan brokers check on applications, and another bank’s mobile work app for its staff.1

On October 7, the security firm CrowdStrike published its investigation. The attacker used ARTEX, an open-source AI penetration-testing tool developed in China, mainly running on DeepSeek’s models, and had also used models from other companies and Claude Code.1

Fact

The investigation could go that far because of the attacker. Their server allowed directory listing, so anyone could read their Claude Code instruction file, session histories, memory files and ARTEX configuration from a public address. From those records, CrowdStrike identified their servers, the models they used and part of what they did.1

Fact

Since then: on October 8, the developer of ARTEX said on GitHub that the tool had been misused and announced that it would stop receiving updates and move to closed source, with no further public releases; the project page has since been taken down. The statement did not refer to the Korean incidents directly.4

Exhibit 1 One person: attacking banks with AI, unable to lock their own server
Attacker One person, with an AI toolkit A free, open-source AI tool that looks for vulnerabilities on its own (ARTEX)
Went after the banks
  1. Looked for weak entry points Systems used by staff, contractors and loan brokers
  2. Peripheral systems breached A loan-progress service for brokers; a mobile work app for bank staff
  3. About 68,000 people’s data exposed Total for the wave of breaches (media count): names, phone numbers, income, loan limits
So: Peripheral systems reach customer data too
Could not lock their own
  1. Their server was left open Anyone could list their files from the internet, no login needed
  2. Their AI work records were exposed The instructions they gave the AI, their conversations with it, tool settings
  3. Investigators pieced it together From those records CrowdStrike identified their servers, the models they used and part of what they did
So: AI work records are sensitive data in their own right
Compiled from sources Source: CrowdStrike (Oct 7), Runtime Wire (Oct 6)

Much about the case is still unsettled. Here is what is established, kept apart from what is not:

Exhibit 2 What is established, and what is not

Established

  • Several South Korean financial institutions were breached between late September and early October CrowdStrike (security firm), Oct 7
  • The breached systems include a loan-progress service for brokers and a mobile work app for bank staff CrowdStrike, Oct 7
  • The attacker used an AI tool that looks for vulnerabilities on its own (ARTEX) CrowdStrike, Oct 7
  • The attacker’s server was left open; anyone could read their AI work records CrowdStrike, Oct 7

Not established

  • How many institutions were affected Media report at least seven; CrowdStrike says the number is unconfirmed
  • Whether every breach was this one person’s work CrowdStrike says the attacker “targeted” these institutions; it does not tie each breach to them
  • How much the AI did on its own Korean media quote the Financial Security Institute: a person used the AI as a tool
  • Who the attacker is Their AI records hold one set of personal details; CrowdStrike cannot confirm they belong to the attacker
  • Whether core banking systems were affected Public reports mention only peripheral systems
Compiled from sources Source: CrowdStrike, The Record, Runtime Wire, The Herald Business, as of October 13, 2026

02The real question

When one login is compromised, how much can an attacker reach?

Most of the commentary stops at how careless the attacker was. But the carelessness is the point: someone who could not secure their own server could mount a wave of attacks on major banks that ended in stolen data. For a company, that contrast highlights two things.

First, an attacker no longer needs to be good at everything. What used to keep many people out was the full set of skills an attack required, and that bar is coming down.

Second, what fell were peripheral systems. A status page for brokers and a mobile app for staff are not what anyone calls critical systems, yet both could reach customer data.

So the real question is this: when a weaker link fails, how much can an attacker reach through it?

03What we think

AI lowers the bar for attackers. What a company can design is the blast radius

Inference

AI tools are pulling attack capability apart from an individual’s skill. Finding targets, probing for weaknesses, reading the errors and adjusting used to take experience; part of that work can now be handed to tools, done faster, by someone who does not know every step. Put as a score: places that once took a 70 to get into may now be within reach of a 40 with AI. South Korea’s president said as much: AI now makes it possible to hack with ease even without specialized skills.2

The case does not show that the AI carried out the intrusion by itself; Korean media quote the Financial Security Institute as saying a person used the AI as a tool.5 But if someone who could not lock their own server could mount attacks like these, that tells you how far the bar has dropped.

Our view

So “is the attacker good enough?” no longer works as a defense. What a company can design instead is how much each account, each system and each AI agent can reach. The smaller that reach, the smaller the damage when one of them fails. Security people call it the blast radius.

Exhibit 3 How much a lost login can reach depends on how much it was given
At an accounting firm: each column is one client (12, for illustration). Suppose the same login leaks.

Everyone shares one login

Books Tax Filings Payroll
Reaches: every record of all 12 clients

Each login covers only its own clients

Books Tax Filings Payroll
Reaches: the records of 4 clients
Data the leaked login can reach What one login (or one AI agent) can reach; outside it, nothing opens
Illustrative Source: Augvant

04Why it matters

At accounting firms, AI agents inherit shared logins

This problem was there before AI.

In many firms the core data still sits in DBF files, where permissions are hard to separate, and the work runs on LINE, email, phone calls, Excel and the memory of senior staff. Two things follow at once. Permissions that should be separate are not, so many people can reach every client’s data. And data that should be easy to find is scattered: people discover it is missing when they need it, and when someone leaves, the client context leaves with them.

This is usually nobody’s fault. With few hands, people covering for one another and tools that cannot draw finer lines, wide permissions and shared folders are how the work gets done. The tools have not kept up.

AI raises the cost. From outside, an attacker with AI finds an unlocked door more easily. From inside, any agent added later inherits the permissions already there: add one to a firm where everyone can reach everything, and you have added one more master key. And one thing is easy to miss: an AI’s work records are sensitive data in their own right. The investigation could reconstruct this attack because the attacker left their AI records unlocked; a company’s agent records will hold customer data too, and they deserve the same lock as the customer files.

What Korean regulators asked of the financial sector afterwards is something a firm can do too: list every system that can be reached from outside, including ones that are not for customers, and add the authentication and access controls that are missing.6 There is no need to wait for AI. Four questions are worth asking now:

  1. Which logins in the firm are shared by several people?
  2. Which clients’ data can each person, and each login, reach? Is all of that needed for the job?
  3. When you need a particular client’s document, do you know where it is?
  4. Where are your AI tools’ records kept, and who can see them?

05What we do about it

Sort out people’s permissions before AI agents take over

Our view

A lower bar for attackers turns boundaries from a job for the security team into a precondition for using AI in a firm at all. Large companies have security teams to set permissions and run audits; an accounting firm usually does not. So the boundaries cannot depend on a separate security product, or on a policy everyone has to remember. They have to be built into the systems the firm uses every day.

Augvant builds operating software for process-heavy small businesses, starting with accounting firms in Taiwan. These are the principles we hold ourselves to.

AI is a means, not the end. What cannot hold up on a real firm’s floor, we do not ship.

See it before you draw the lines. How much a person should reach depends on how the work actually flows. So inside a firm, our first step is not code. It is making the operation visible: how data flows, where money moves, which deadlines are coming due, where work is stuck.

Get people’s permissions right before agents take over anything. Once it is visible, we steady the most fragile points first, then redesign the process and the permissions: each person reaches the clients they look after, and data has a fixed place where it can be found. Agents join on that footing. Each has a role, a scope and someone who answers for it, and its scope is never wider than that person’s; what falls outside it goes to a person.

Every step has to be traceable. DBF is a file format more than forty years old, not a proper database, and who changed what is hard to trace. The bookkeeping system we are developing lets firms move off their old files onto a proper foundation for data permissions and security; AI can operate it, and every step can be looked up afterwards.

In this case, the attacker could borrow their capability from AI, but no one was going to guard their boundaries for them. Firms are no different: capability can be borrowed; boundaries have to be designed. What we want is for firms without a security team to have well-designed boundaries all the same.

Sources

  1. CrowdStrike, “Unknown Threat Actor Uses AI-Driven ARTEX to Target South Korean Finance,” October 7, 2026. https://www.crowdstrike.com/en-us/blog/unknown-threat-actor-uses-artex-to-target-south-korean-finance/ ↩ ↩2 ↩3 ↩4

  2. Suzanne Smalley, “South Korean officials believe AI agents were used to hack several banks,” The Record, October 6, 2026. https://therecord.media/south-korean-bank-hacks-ai-agents ↩ ↩2

  3. Ryan Merket, “AhnLab finds ARTEX on 600 IPs as police probe South Korean bank breaches,” Runtime Wire, October 6, 2026. https://runtimewire.com/article/south-korea-bank-breaches-artex-ai-probe ↩

  4. Kim Byeo-ree, “Developer of AI hacking tool used in South Korean financial breaches to go private,” The Herald Business, October 10, 2026. https://biz.heraldcorp.com/article/10898148 ↩

  5. Jeong Ho-won, “Exclusive: Chinese AI tool ARTEX used in wave of bank hacks, probe finds,” The Herald Business, October 3, 2026. https://biz.heraldcorp.com/article/10892497 ↩

  6. Bill Toulas, “South Korea probes bank breaches amid suspected AI-powered attacks,” BleepingComputer, October 5, 2026. https://bleepingcomputer.com/news/security/south-korea-probes-bank-breaches-amid-suspected-ai-powered-attacks ↩

Questions, or a different reading of the same facts? We’d like to hear it.

Email us