The ARTEX AI Claude agents cyberattack on South Korean banks, and why it is not only an AI story
The ARTEX AI Claude agents cyberattack on South Korean banks landed on Shinhan Bank, KB Kookmin Bank, and Hana Bank earlier this month, according to BleepingComputer’s report on CrowdStrike’s findings (opens in new tab). A Chinese-speaking operator used an agentic penetration testing suite called ARTEX AI, driving Claude Code sessions through it, to reach customer personal data and card data. Some systems went down. The Korean government called an emergency meeting and pushed immediate security directives at operators of critical IT systems.
What makes this different from the breach writeups I read every other week is the tooling. ARTEX is not a bespoke exploit chain someone spent months building. It is a suite that, until recently, anyone could clone. The attacker supplied targets and prompts. The model supplied the reasoning and the command generation. That collapse in effort is the actual story, and it has less to do with AI capability than with what a capable operator can rent instead of build.
I have watched the cost of a foothold fall for years. Scripted scanners gave way to frameworks, frameworks gave way to managed kits, and now the expensive part is knowing which system to point at. ARTEX did not invent that problem. It made the problem cheaper.
I want to be careful about sourcing here, because the identity claims in this story are the shakiest part. CrowdStrike confirmed ARTEX usage. CrowdStrike also found a résumé the attacker built with the same AI tools, which leaked name details, contact information, and a Telegram handle. That résumé lists a 2007 birth date alongside a profile that otherwise reads as a 26-year-old who studied at the South China University of Technology. Researchers declined to confirm identity off it, which is the right call. Tool attribution and person attribution are not the same confidence level, and the file that looks most damning is the one I trust least.
The rest of this post covers the LLM stack behind ARTEX, the open directories that gave investigators their leads, the Korean government response, and what any team running agentic pentest tooling should actually do about it.
Background: ARTEX AI, Claude Code, and how the toolchain fits together
ARTEX AI began life as an open-source agentic penetration testing suite built in China, and CrowdStrike confirmed it was used in these intrusions, which is the fact that anchors everything else in the reporting. Agentic here means the tool plans steps, calls an LLM to decide what to run next, and executes it, rather than waiting for a human to paste each command. Claude Code was the operator’s shell in that loop. The attacker typed objectives; the model wrote and ran the commands.
CrowdStrike’s account of the backend is specific. The ARTEX instance used DeepSeek v4.1-flash as the primary LLM, supplemented with GLM-5.3 from Zhipu AI and Grok 4.6 for additional Claude Code sessions. That mix is worth sitting with for a moment, because it tells you the operator treated models as interchangeable capacity rather than as a single trusted vendor. When one endpoint gets rate limited or refuses a request, you route around it.
The access path CrowdStrike points to runs through xcai[.]pro, described as a likely LLM API proxy and reseller. This is the part of the story I find most interesting, and the part that is most mundane. Resold API access is a grey market. Someone buys credits, wraps an endpoint, and sells inference to anyone with a payment method, sometimes with the origin provider’s terms quietly broken. Attribution evidence tends to die at that layer, because the account behind the traffic may belong to a reseller who never asked a question. It is roughly the same problem as a hosting provider who resells cPanel licences through three tiers of resellers and cannot tell you who actually owns the box abusing your mail server. The paper trail exists. It just does not lead to a name.
Investigators also found open directories on the attacker’s own infrastructure. Those directories held Claude Code session histories, ARTEX configuration files, and Claude memory files, all readable. CrowdStrike’s findings on the tooling are covered in BleepingComputer’s report (opens in new tab), and the session histories in particular are what let researchers reconstruct intent rather than guess at it.
What’s happening now: CrowdStrike links the ARTEX AI penetration testing suite to Shinhan, KB Kookmin, and Hana
CrowdStrike’s confirmation that ARTEX AI was in play matters less on its own than what came with it. The target list in these intrusions overlaps with earlier breach reporting against the financial sector, which is what let researchers move from “an agentic pentest suite was used somewhere” to linking this activity with high confidence to prior campaigns. That overlap is the load-bearing evidence, not the tooling. Tools get copied. A consistent set of targets, infrastructure, and session artefacts is harder to fake.
The Korean government did not treat it as a single-bank incident. An emergency meeting followed, along with immediate security directives for critical IT systems. That response is proportionate if you consider what actually leaked: personal data and credit card information from customers at Shinhan Bank, KB Kookmin Bank, and Hana Bank, plus outages in some systems. When card data is in scope, the remediation work runs far past rotating a password. Card reissuance, notification obligations, and regulator scrutiny all land on the banks at once.
The exposed files are where the story gets strange. The same session histories that gave researchers their reconstruction also show the operator asking Claude to suggest Telegram groups that sell data focused on Korea. If you were running a professional extortion operation, that step would sit inside a plan that had already priced the data, lined up buyers, and worked out how to handle the payout. What is on file instead looks like someone who had broken in, taken what was there, and was working out what to do with it afterwards. CrowdStrike’s records indicate no specific monetization plan.
Then there is the résumé. The operator used the same AI tooling to build one, and left the output where it could be read. Identification details, contact information, and a Telegram account came out with it, which is roughly the digital equivalent of burgling a shop and leaving your CV on the counter. I have cleaned up after messy incidents on my own boxes, including a forgotten staging VPS with directory listing enabled, and the pattern is always the same: the operational mistakes are boring and repetitive, not clever.
I would not read too much into the résumé yet, though. It is a lead, and a strong one, but the document is a piece of self-presentation produced with a model that will happily write whatever it is asked to write. Treat it as direction, not identification.
What it means in practice for teams running agentic pentest tools
The most useful artefact to come out of this case is not a malware sample. It is a directory listing. Open directories on the operator’s own infrastructure served up Claude Code session histories, ARTEX configuration files, and Claude memory files, and those files did more for attribution than anything pulled off a compromised host. There is an uncomfortable mirror in that for anyone who runs infrastructure. I once found an old staging VPS of mine with autoindex on sitting in the nginx config, quietly publishing a directory that held deploy keys. Nobody had to break into it. It was handing things over.
The ARTEX maintainer’s response was to close the source and stop shipping updates. That decision only reaches so far. The code was already public before the change, and English and Korean derivatives are circulating, translated and repackaged by people who never needed the original author’s permission. Closing the upstream repository is like pulling your own download page while the file sits on a few hundred mirrors.
Attribution is where the reporting gets thin. The résumé points to a 26-year-old who studied at the South China University of Technology and lives in Maoming, Guangdong, but the same record carries a 2007 date of birth, which does not fit. CrowdStrike says the personal details are not reliable enough to confirm identity, and that caution is the right call.
The part that should worry defenders is the unremarkable part. No zero-day appears anywhere in this chain. An agentic suite driving commercial models through a resold API was enough to reach customer data at three banks. If your threat model assumes the expensive intrusions require expensive exploits, this case does not support that assumption.
Does the ARTEX AI penetration testing suite still pose a risk now that it is closed source?
Yes. The repository was never the dangerous part, the capability was, and the capability has already been copied. Derivatives exist in usable form, so the upstream decision changed who maintains the code rather than who can run it.
Are Claude agents themselves the vulnerability here?
No. Claude Code was the operator’s shell. The actual failures were exposed directories, resold API access with no accountability behind it, and ordinary control gaps on the bank side. Blaming the model is the least useful conclusion on offer.
Does the leaked Claude Code session history prove the attacker’s identity?
Not on its own. It links the activity with high confidence and describes the person behind the workstation, but the contradictory birth date is precisely why researchers stopped short of naming anyone.
What to expect next for the ARTEX AI Claude agents cyberattack on South Korean banks
I expect the derivative count to rise, not fall. Closing an upstream repository works when the code lives in one place and everyone depends on the maintainer’s releases. That is not the situation here. The ARTEX codebase sat in public long enough to be forked, translated into English and Korean, and reworked by people with no relationship to the original developer. Those forks do not stop compiling because the maintainer stopped pushing commits.
The overlap between the Korean bank targets and earlier financial-sector breach reporting is the detail I keep returning to. An actor who already had finance-sector reconnaissance on file was not running a single-country operation. My expectation is that regulated targets in other regions will see the same pattern, the same dependence on resold model access, and that most of them will find out when a report names them rather than from their own telemetry.
That points at the layer almost nobody audits. API resellers and proxies such as xcai[.]pro sit between a paying customer and a model provider, and they dissolve the accountability providers try to build into their platforms. Investigators keep finding attribution evidence at that tier because a paper trail has to exist somewhere between the card payment and the inference request. I would expect more enforcement pressure there, and I would expect it to move slowly, since reselling access across borders is not one company’s problem to fix.
On identity, expect the profile to keep wobbling. A résumé generated with the same AI tools the operator used for everything else is a lead, and the 2007 date of birth is enough of a contradiction to keep it filed as one. Watch for Korean authorities to publish something that either corroborates the Maoming profile or quietly drops it. The absence of a follow-up will be its own signal.
The one outcome I would not bet on is a fix at the model layer. Anthropic did not ship this capability as a product feature, and no provider will make an agentic pentest suite stop being useful by tightening a usage policy. The capability lives in the surrounding infrastructure, and so does the leverage to slow it down.
Your next step
The exposure that gave CrowdStrike its best material was an open directory, and it is worth checking whether you have one before you do anything else on this list. On nginx it is a single line, autoindex on;, often added to a vhost during a file-sharing test and left in place. On Apache it is Options +Indexes inside a directory block. Both produce a browsable listing of everything underneath, and neither will show up in a normal uptime check. I have found this exact misconfiguration on a staging box of my own that was three years old, serving a directory of customer invoice PDFs to anyone who happened to know the IP. Nobody had touched that vhost since the site launched.
A pass looks like this. Enumerate every public IP in your inventory, including the ones sitting in DNS records you forgot about, and request a directory path on each that should not be listable. If you get a file index back, you have the same problem the ARTEX operator handed to investigators, minus the luck of it being someone else’s infrastructure.
Then go through your model access. If anyone on your team is calling a resold or proxied endpoint rather than a first-party API, rotate those keys today and find out who added the proxy and why. The reason is not that the proxy is necessarily hostile. It is that you cannot answer basic questions about where the prompts went, who retained them, or what happens to your keys if the reseller is compromised. Providers ship key scoping, spend limits, and audit logging for a reason, and a proxy sits in front of all three.
The third item is smaller and easier to postpone. Write down what your detection would actually catch if someone pointed an agentic pentest suite at your perimeter tomorrow. That means the specific alert, the specific log source, the person who gets paged, not the policy document that says it will not happen. If you cannot name a single alert, that gap is what you take to whoever owns your security budget this quarter, and it will be a shorter conversation than a compliance summary.
Start with the directory scan. It takes an afternoon, needs no new tooling, and it is the one item here that sits entirely inside your control.