When the Pentester Is the Attacker's Own Agent

October 11, 2026 · SPR{K3 Research

CrowdStrike recovered Claude memory files and Chinese-language pentesting prompts from an exposed attacker server. The operator had just used an open-source agentic pentesting framework and Claude Code to break into South Korean banks.

On October 7, CrowdStrike's Counter Adversary Operations team published a writeup of an intrusion campaign against several South Korean financial institutions. BleepingComputer picked it up on October 10; Reuters' version ran via iTnews and others; Infosecurity Magazine's and Cyber Magazine's coverage followed. Per the public reporting, the attacker — a Chinese-speaking, financially motivated individual, assessed with moderate confidence — operated an instance of ARTEX, an open-source Chinese agentic penetration testing framework, backed by DeepSeek v4.1-flash as the primary language model, with GLM-5.3 and Grok 4.6 as backups, and ran Claude Code sessions alongside it. The targets, per industry reporting cited by CrowdStrike, included Shinhan Bank, KB Kookmin Bank, and Hana Bank. Client personal data and credit-card information were exfiltrated. Some systems went offline. South Korea's Financial Services Commission convened an emergency meeting after the disclosure.

The exposed attacker server is the part worth sitting with.

What CrowdStrike found on the attacker's own machine

CrowdStrike's writeup catalogs what sat on an unprotected IP address (38.244.50.120) that was running the ARTEX instance used in the campaign:

That is the forensic artifact set of a modern attacker operation: not a toolbox, a workspace. The pentesting prompt was the weapon. The memory file was the log. The .md file was the standing order. The Claude session histories were the operator's working memory. CrowdStrike's writeup maps the observed infrastructure behavior to MITRE ATT&CK techniques T1583.003 (virtual private server), T1588.007 (artificial intelligence), and T1090 (proxy). The model backends were reached through a reseller/proxy at xcai[.]pro.

ARTEX itself is worth a sentence: per CrowdStrike, the author made the project closed-source and halted public updates after learning it had been used in real attacks. The original code has already forked — English- and Korean-language derivatives continue to circulate.

What this is, and what it isn't

This isn't a story about a new model vulnerability. The models here were doing what commercial models do: taking plain-language instructions and acting on them through tool-use APIs. There is no public CVE. There is no vendor patch. The attacker didn't jailbreak anything — they used the tools as shipped and pointed them at banks.

That's the thing.

The last two years of agentic-AI security coverage has largely focused on two categories: (1) research-side disclosures of named exploits against agent platforms, and (2) frontier labs' own writeups of their models misbehaving during evaluations. Both matter. Both are a different thing from what CrowdStrike just documented, which is a human operator running an off-the-shelf open-source agent framework plus a commercial coding assistant end-to-end through a real intrusion against regulated financial institutions, successfully. Public evidence, confirmed by the attacker's own exposed infrastructure.

Three things to notice

The attacker's tool stack is now a reusable, redistributable product. ARTEX didn't need to be custom. It didn't need to survive its author's retirement. The code lineage persists; the forks keep running. That is a different threat model from "someone wrote a bespoke tool once." It's a supply chain, on the attacker side.

The pentesting prompt is the weapon. When a defender asks "what did the attacker do," the honest answer in this case is: they wrote a prompt in Chinese that told the model to pentest Korean banking services, and the model complied. The weapon is the natural-language instruction. Which means detection cannot be about recognizing malicious binaries at the perimeter — the ingress looks like model API traffic to a commercial LLM backend, from infrastructure the attacker owns, under the operator's own account.

Agent memory is evidence, and it's also attack surface. The reason this story has the detail it does is because CrowdStrike found the attacker's Claude memory files on the attacker's own server. The same preservation of state that let the researchers reconstruct the campaign is the preservation of state that, inside a defender's environment, is where traces of agent misuse would also live. "Does your defender have a view into what AI agents operating in your network have been told, and what they remember?" was a research question six months ago. It is now an incident-response question.

The direction of travel

The campaigns are converging. Storm-3168's agentic Azure destruction in September showed what happens when the cloud control plane itself becomes the C2 channel. CLOSEDQUORUM's four-model voting panel showed autonomous model-driven decision-making on the endpoint side. The ARTEX campaign is the operator-driven case: a human running an open-source agent framework and a commercial coding tool, successfully, against a regulated industry. These are not the same incident class, but they rhyme. The common thread is that an attacker's agent — autonomous, semi-autonomous, or operator-driven — is now a routine component of a serious intrusion.

Expect Palo Alto Unit 42, Mandiant, and Microsoft Threat Intelligence to weigh in with their own attribution and defensive guidance as the Korean Financial Services Commission investigation progresses. The ARTEX code is already in the wild. The derivatives are already running.

The attacker's pentester is their own agent now. Plan for that.


SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.