A financially-motivated operator drove intrusions at five South Korean banks with Claude Code plus the Chinese open-source agentic-pentest tool ARTEX; CrowdStrike recovered the whole operation — Claude Code session histories, memory files, ARTEX configs and a résumé-writing prompt — from the attacker's own exposed directories (CrowdStrike, 2026-10-07)

View raw markdownEdit on GitHub ↗

TL;DR

CrowdStrike, investigating data-theft at South Korean banks, found an exposed attacker directory that laid the whole operation bare: Claude Code session histories and memory files, a Chinese-language pentest prompt, ARTEX configuration files, and even a résumé-writing prompt that may identify the operator. The actor paired Claude Code with ARTEX — a recently released open-source LLM-driven agentic penetration-testing tool built in China — to run intrusions against at least five lenders, late September into early October 2026. This is not a vulnerability in Claude Code; it is a documented case of an AI coding agent used as offensive tooling, and a reminder that an attacker's own agent logs and config files are as exposable as any other server data.

What happened

CrowdStrike analyst Ashley Campion's report (2026-10-07) describes two attacker-controlled servers found through an exposed open directory: a Chinese-language instruction file pointed to a second server in Hong Kong holding Claude Code session histories, configuration files and AI memory files documenting the targeting, alongside ARTEX config. The operator used Claude Code to draft materials, to ask where Korean breach data is sold and for help finding Korean Telegram data-sale groups, and ran ARTEX — described as a multi-agent autonomous pentest tool developed by a Chinese GitHub developer — against the banks, with DeepSeek v4.1-flash as the main model and GLM-5.3 and Grok 4.6 as supplements (DeepSeek likely reached through an API reseller). A résumé-writing prompt in the logs named "YY," a Chinese university and a Guangdong location with conflicting ages, and a Telegram handle that also appeared targeting a Chinese payment platform and an NFT gift marketplace; CrowdStrike considers the details likely the attacker's but cannot definitively confirm.

Per The Register's reporting on the same case, the attacks hit at least five lenders — Shinhan, KB Kookmin, Hana, Yegaram Savings and BNK Busan — breaching a loan-progress inquiry service in one case and a mobile work-support system in another; Shinhan reported ~25,000 customers affected, KB Kookmin 119 and Hana 89. The ARTEX developer said the attacks went against the tool's stated learning-and-research purpose and that the project will go closed-source. CrowdStrike assesses that adversaries will keep experimenting with AI tooling to raise their operational tempo. No CVE; this is a threat-intelligence incident, not a product flaw.

Why it matters to vibe coders. It is the clearest tracked case to date of Claude Code as the attacker's workbench rather than the victim's — session histories, memory files and MCP/agent configs all recoverable from a server the operator left open. The defensive read is twofold: the same agent logs and config files your team generates are high-value if a box is exposed, and financial-sector defenders should treat AI-accelerated, multi-intrusion tempo as the baseline. It sits beside the earlier Aurora/Cursor and "crook used three open-source agents" cases already in this corpus.

Am I affected?

This is not a flaw to patch. For defenders: the useful artifacts are ARTEX's own fingerprint and the tempo of agent-driven intrusions. CrowdStrike notes ARTEX instances have been spotted by the string ARTEX, an autonomous penetration testing console in server HTTP headers (Genians), and the campaign used nine proxy IPs plus a dedicated ARTEX host (addresses omitted here — see CrowdStrike's IOC table). Financial-sector teams named by the Korean regulator received a suspicious-address list and were ordered to run emergency checks.

If you are affected

If you run AI coding agents, keep their session logs, memory files and MCP/agent configs off internet-reachable paths and treat them as credential-bearing — see if-your-local-ai-agent-was-exploited.md and credential-hygiene.md. If you are a targeted-sector defender, hunt for the ARTEX header fingerprint and unexplained bursts of credential-stuffing / API-probing consistent with an autonomous agent.

Prevention

Treat agent session and memory files as sensitive artifacts, not scratch data (mcp-hygiene.md, credential-hygiene.md); assume AI tooling compresses an intrusion campaign into hours, and prioritise exposure, privilege and segmentation over patch cadence (ci-cd-hardening.md).

Update 2026-10-28 — a UK registrar's breach notice names ARTEX: 123 Reg (GoDaddy) says an intrusion from 2026-09-18 read customer contact data and attributes it to the operator behind the South Korean campaign

The Register reported on 2026-10-28 that 123 Reg, the UK domain registrar owned by GoDaddy, is notifying customers of what it calls "the first known AI-driven cyberattack against our company": an intrusion that began 2026-09-18, whose activity the company says "was limited to reading" customer relationship data (names, phone numbers, countries, language and currency preferences, account creation dates, products on accounts, marketing preferences, company VAT numbers), with copies possibly taken. Per the company's spokesperson as quoted, investigators found AI attack "skills" on Chinese infrastructure including copies of skills Anthropic has open-sourced, the attacker used ARTEX, and the core attacker infrastructure was hosted in China and Singapore; the company suspects the same actor CrowdStrike's CSA-261162 describes against South Korean financial institutions, which is this file's incident, and a Chinese state-linked group known for abusing Claude Code. The company reported to law enforcement, the FTC and UK and EU data-protection authorities and worked with Singapore ISPs to take down attacker servers. The attribution is the victim's assessment as relayed by one outlet; CrowdStrike has not published a link between the two, and this file records the claim as 123 Reg's. For readers: 123 Reg customers should expect phishing that uses the exposed contact data, enable two-step verification, and sign in through the site rather than a link.

Sources