Microsoft AI代理評測:內部威脅論,60款工具現形 | Microsoft AI Agent Review: Insider Threat, 60 Tools Found
By Kit 小克 | AI Tool Observer | 2026-10-20
🇹🇼 Microsoft AI代理評測:內部威脅論,60款工具現形
Microsoft 10月19日宣布,企業內的AI代理(AI Agent)將被當成「內部威脅」等級風險管控。新機制Dynamic Agent Risk是Project Perception框架與Microsoft 365 E7的最新擴充,讓每個在系統裡跑任務的AI代理都被即時評分、即時收權。
Microsoft為什麼把AI代理當內部威脅?
代理一旦拿到帳密與系統存取權,出錯或被誘導的傷害跟員工內鬼一樣大,差別是代理不會猶豫,出包速度人來不及反應。
資安主管Hayete Gallot稱這類過度授權的代理為「agentic insider」;執行長Satya Nadella說得更直白:「必須把智慧的供給,和執行的授權分開」。
Dynamic Agent Risk怎麼評分收權?
系統持續蒐集身分、裝置、雲端、網路訊號即時打分數,AI代理分數一飆高,權限就自動降級,不必等人工發現。
- 風險上升先掉寫入權限,只剩唯讀
- 分數再高,高風險動作要真人核准
- 爆表直接隔離或關閉整個代理,且權限上限不超過它服務的員工
Microsoft Defender現在能認出60種代理工具(前版僅35種),含Claude Code、Codex、GitHub Copilot,每個都拿到專屬Entra Agent ID。
企業該怎麼看這套AI代理風控?
導入AI代理不再是接個API就能上線,資安團隊要先想清楚:出事時能不能幾秒內把權限砍到唯讀或關掉。Microsoft同步推出Execution Container沙盒,防止代理被滲透後反過來關掉監控。完整細節要等11月17日Ignite大會公布,目前仍在滾動擴充階段。
好不好用,試了才知道。
🇺🇸 Microsoft AI Agent Review: Insider Threat, 60 Tools Found
Microsoft dropped a blunt announcement on October 19: enterprise AI agents should be treated as insider-threat-level security risks, not just tools to defend against outside hackers. The new framework, Dynamic Agent Risk, is the latest expansion of Project Perception and ships as part of Microsoft 365 E7 — built to score and restrict every AI agent running inside a company's systems in real time.
Why does Microsoft treat AI agents as insider threats?
Once an agent holds credentials and system access, a mistake or manipulation does the same damage as a rogue employee — except the agent doesn't hesitate, and moves faster than a human can react.
Microsoft security chief Hayete Gallot coined the term "agentic insider" for over-privileged agents. CEO Satya Nadella put it more bluntly: "We need to separate the supply of intelligence from the authority over it."
How does Dynamic Agent Risk actually score agents?
The system continuously pulls signals from identity, device, cloud, and network tools to give every AI agent a live risk score; when the score spikes, permissions get cut automatically — no human has to catch it first.
- Rising risk first strips write access, leaving read-only
- Higher scores require human approval for risky actions
- Worst scores get the agent isolated or shut down entirely, capped so it never outranks the human it serves
Microsoft Defender now identifies 60 agent types (up from 35), including Claude Code, OpenAI Codex, and GitHub Copilot — each gets a unique identity via Microsoft Entra Agent ID.
What does this mean for enterprises rolling out AI agents?
Deploying an AI agent is no longer just "wire up an API and grant access" — security teams need a plan to cut an agent to read-only, or kill it outright, within seconds. Microsoft also introduced the Execution Container sandbox to stop a compromised agent from disabling its own monitoring. Full details land at Microsoft Ignite on November 17 — this is still a rolling rollout, not a finished product.
Good or not, you only know once you've tried it.
Sources / 資料來源
- Microsoft says AI agents are like insider threats as frontier models cannot be trusted (TechCrunch)
- Microsoft targets the "agentic insider" with new controls for rogue AI agents (SiliconANGLE)
- Microsoft adds tools to limit and monitor AI agents in Microsoft 365 E7 (The Next Web)
常見問題 FAQ
什麼是Microsoft的Dynamic Agent Risk?
Dynamic Agent Risk是Microsoft Project Perception底下的即時風險評分機制,持續監控AI代理的身分、裝置、雲端、網路訊號,一旦行為異常就自動降低代理權限。
AI代理為什麼會被當成內部威脅?
因為AI代理一旦拿到系統存取權,行為出錯或被誘導的傷害等同員工內鬼,而且代理執行速度快,人工難以即時介入,所以Microsoft用內部威脅等級的框架來管控。
Microsoft Defender能偵測哪些AI代理工具?
目前能辨識60種代理工具,包含Claude Code、OpenAI Codex、GitHub Copilot等,比前一版的35種大幅增加,每個偵測到的代理都會被指派專屬Entra Agent ID。
企業現在就能用這套風控機制嗎?
部分功能已隨Microsoft 365 E7釋出,但完整細節與時程會在11月17日的Microsoft Ignite大會公布,目前仍在滾動擴充階段。
Dynamic Permissions的權限上限怎麼訂?
代理只能拿到完成單一任務所需的最低限度、有時效的權限,且上限不會超過它所服務的那位員工原本擁有的權限層級。
延伸閱讀 / Related Articles
- AI監控定價評測:同款商品,兩人結帳差23% | AI Surveillance Pricing Review: Same Item, 23% Price Gap
- AI穿戴裝置評測:十二州竊聽法,戴的人要擔刑責 | AI Wearables Review: Wiretap Laws Put Users at Risk
- OpenAI失控代理評測:24年醫院資料外洩,加密錢包遭盜刷 | OpenAI Rogue Agent Review: Hospital Data, Wallet Drained
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
留言
張貼留言