ARTEX評測：AI滲透測試工具淪為駭韓銀行武器 | ARTEX Review: AI Pentest Tool Weaponized in Bank Hacks

10月 12, 2026

By Kit 小克 | AI Tool Observer | 2026-10-12

🇹🇼 ARTEX評測：AI滲透測試工具淪為駭韓銀行武器

這週資安圈討論最熱的不是哪家公司發新模型，而是開源AI滲透測試工具 ARTEX 被證實淪為真實攻擊武器——駭客用這套原本設計給資安團隊練習的工具，入侵至少七間南韓金融機構，外洩約6.8萬人個資。這起事件，把「AI agent 滲透測試」從概念驗證拉進了真實資安事故名單。

ARTEX 是什麼：LLM 驅動的自動化滲透測試系統

ARTEX 由開發者 Autumn-27 於2026年7月26日在 GitHub 以 AGPL-3.0 授權釋出，曾在百度安全應變中心主辦的「Agent+」競賽中，從約150支隊伍脫穎而出奪冠。它讓多個 LLM agent 協作：

自動掃描目標、找出漏洞，不需人工逐步下指令
自行規劃攻擊路徑並呼叫既有資安工具執行
驗證攻擊結果，後端採 Go 搭配 Next.js 前端與 PostgreSQL

換句話說，ARTEX 把紅隊演練的「判斷—執行—驗證」迴圈交給 AI agent 自主完成——這正是它受關注的原因，也是它變危險的原因。

真實攻擊：南韓銀行遭入侵，6.8萬人個資外洩

CrowdStrike 追查發現，2026年9月底到10月初，有攻擊者利用ARTEX入侵新韓銀行、Yegaram 儲蓄銀行等至少七至九家南韓金融機構，竊取姓名、電話、年收入與貸款額度等資料，約6.8萬人受影響。從攻擊提示詞多為中文研判，幕後較可能是財務動機的中文使用者，而非國家級行動。

事件曝光後，開發者 Autumn-27 於10月8日宣布將原始碼轉為私有，並停止後續版本發布與維護——等於直接把工具下架。

給導入 AI 滲透測試工具前的提醒

限制部署環境，別讓測試系統直通正式網路
留意開源專案授權與維護狀態，說關就能關
完整留存 AI agent 操作紀錄，方便事後稽核

好不好用，試了才知道。

🇺🇸 ARTEX Review: AI Pentest Tool Weaponized in Bank Hacks

The AI security story making the rounds this week is not a new model release — it is ARTEX, an open-source AI pentesting tool that got weaponized in real attacks against South Korean banks. It is a clean case study in what happens when autonomous AI agents meet dual-use security tooling.

What ARTEX Actually Is

ARTEX is an LLM-driven, multi-agent autonomous penetration testing framework, published on GitHub on July 26, 2026 by a developer going by Autumn-27, under an AGPL-3.0 license. The project placed first out of roughly 150 teams in the Agent+ challenge hosted by Baidu Security Response Center. Under the hood:

Multiple LLM agents scan targets and find vulnerabilities without step-by-step human commands
The system plans an attack path and calls existing security tools to execute it
Results get verified automatically, on a Go backend with an embedded Next.js frontend and PostgreSQL

In short, ARTEX automates the find-exploit-verify loop that used to require a skilled human red-teamer — exactly what made it useful, and exactly what made it dangerous.

The Real-World Breach: 68,000 Records Exposed

CrowdStrike traced a late-September to early-October 2026 intrusion campaign against at least seven to nine South Korean financial firms, including Shinhan Bank and Yegaram Savings Bank, back to an operator running ARTEX. Names, phone numbers, annual income, and loan limits were exposed for roughly 68,000 people. Chinese-language prompts found in the attack workflow point to a financially motivated actor rather than a state campaign.

After the attribution went public, Autumn-27 announced on October 8 that the source would go private, with no further releases or maintenance support — effectively pulling the tool rather than patching it.

What This Means If You Are Evaluating AI Pentest Tools

Never let an autonomous pentest agent touch production networks directly
Review license terms and maintenance status before adopting a tool — open-source security tools can vanish overnight
Keep a full audit log of everything an AI agent actually executes

好不好用，試了才知道 — you will not know if a tool is good until you try it. In this case, someone already did, and 68,000 people paid the price.

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士