Taiwan Government Attack Used Near-Autonomous AI Agents, Report Says
iThome reported that foreign hackers used an AI attack system built on Hermes and OpenClaw agent frameworks in a hybrid attack against Taiwan government agencies in early July 2026. Taiwan’s Ministry of Digital Development said government cybersecurity monitoring detected the anomalous attack in July and that the attackers combined AI agents with other methods. According to Dream Security’s reconstruction from 1,395 work files, the operation lasted less than four days and involved 12 attack waves, with up to eight agents deployed in a single wave. The company said the system cracked 85 sets of government personnel credentials and obtained at least 2,564 personnel records. Dream Security described the operation as near-autonomous, meaning it was not fully autonomous or necessarily free of human involvement. The attacker’s identity and state affiliation were not established.
The report describes a real-world operation in which multiple agents performed reconnaissance, credential attacks, vulnerability validation, research, and strategy adjustment. Dream Security said the intrusion relied on established server-side weaknesses, including unauthenticated APIs, and on SSO trust relationships that enabled lateral movement across systems.
如果駭客攻擊可以同時派出多個AI Agent分頭行動，進行目標偵察、攻擊路徑分析，還能研究新的攻擊方法、自動修正錯誤判斷，並且持續動態調整攻擊策略，這樣的駭客攻擊行動，還能只稱為自動化攻擊嗎？
2026年7月初，就發生了這麼一場針對臺灣政府機關的網路攻擊。雖然攻擊行動不到4天，但它明確顯示AI自主攻擊已經不再只是概念驗證，而開始進入真實世界的網路攻擊行動，成為現在進行式。
這起事件的曝光，是資安公司Dream Security在8月12日揭露一份AI自主攻擊行動的完整工作資料，他們從1,395個工作檔案還原出一場為期約4天、歷經12波攻擊行動的AI網攻事件，並且首度披露單一波攻擊行動部署多達8個AI Agent同時執行任務，最終破解85組政府人員帳密，取得至少2,564筆人員資料，而且還將攻擊行動範圍擴大到與政府機關有關的資訊供應商、核能安全機構及能源業者。
#Dream Security#OpenClaw#Hermes#Taiwan Ministry of Digital Development#Tesseract#Keycloak