跳到主要內容

GenAI.mil評測:五角大廈開放ChatGPT、Grok給300萬人 | GenAI.mil Review: Pentagon Opens ChatGPT, Grok to 3M

By Kit 小克 | AI Tool Observer | 2026-09-02

🇹🇼 GenAI.mil評測:五角大廈開放ChatGPT、Grok給300萬人

GenAI.mil 是美國國防部的內部AI入口,8月31日一口氣新增了 ChatGPT MilGrok for Government 兩個選項,加上原本就有的 Google Gemini,讓國防部300萬名軍職與文職人員可以在通過安全認證的環境裡使用商用大型語言模型,不必把敏感資料丟進一般消費版 ChatGPT。目前已有超過170萬人開通帳號,這是美軍史上規模最大的一次AI採用行動,也讓「政府用什麼AI」變成這幾天全球科技圈討論最熱的話題。

GenAI.mil是什麼?三選一的安全AI入口

GenAI.mil的定位很單純:把民間最強的AI模型「搬進」一個符合國防部資安規範的環境。ChatGPT Mil已通過 Impact Level 5(IL5) 認證,這是美國政府對敏感未分類資訊(CUI)的雲端安全門檻,支援文件、專案與客製化GPT,用來處理規劃、政策、後勤與行政等「文書密集型」工作。Grok for Government則主打推理能力,提供自動、快速、專家三種思考模式,官方說法是要協助「從採購分析到供應鏈管理」的軍事場景。

為什麼Claude不在名單裡?

眼尖的人會發現,三大模型商中唯獨少了Anthropic的Claude。原因不是技術問題,而是一場還在打的官司:國防部今年3月把Anthropic列為「供應鏈風險」,理由是Anthropic拒絕全面開放Claude給國防部使用於自主武器與國內大規模監控等用途,堅持保留安全防護機制。這件事被聯邦法官在8月底判定「非法且毫無根據」,法院認定國防部的說法「完全站不住腳」,但Anthropic目前仍被排除在國防部合約之外,官司還在打。換句話說,GenAI.mil的名單反映的不只是技術優劣,更是各家公司在「安全防護要留多少彈性給客戶」這件事上的立場差異。

Grok的資安疑慮:CSAM問題「無解」仍照常上線

更值得注意的是,根據多家外媒引述的內部資訊,xAI工程師曾表示Grok生成兒少性剝削內容(CSAM)的風險目前沒有可靠的技術解方——因為能生成成人露骨內容的模型,本質上也具備把提示詞裡的「對象」換成未成年人的能力。即便如此,Grok for Government仍在沒有公開風險揭露的情況下,取得IL5臨時授權並部署到300萬名國防部人員的平台上。xAI目前握有國防部2億美元合約。

給企業與開發者的啟示

  • 政府採用AI≠安全驗證完成:IL5認證管的是資料儲存與傳輸安全,不等於內容生成風險已被排除
  • 「安全防護」正在變成商業籌碼:願意放寬限制的廠商可能更快拿到大型政府合約,但這未必代表產品更成熟
  • 多模型並存是趨勢:GenAI.mil同時上架三家模型,說明大型機構已經不押寶單一供應商

這起事件的重點不在於選哪家AI,而在於「AI政府採用」背後那套看不見的安全與法律角力,才是決定哪個模型能上場的真正關鍵。好不好用,試了才知道。


🇺🇸 GenAI.mil Review: Pentagon Opens ChatGPT, Grok to 3M

GenAI.mil, the Pentagon's internal AI portal, added ChatGPT Mil and Grok for Government on August 31, joining the existing Google Gemini deployment. The move gives 3 million Department of Defense military and civilian personnel access to commercial large language models inside an accredited security environment, instead of routing sensitive work through consumer-grade chatbots. Over 1.7 million users are already onboarded, making this the largest AI rollout in U.S. military history — and the story dominating AI news cycles this week.

What Is GenAI.mil? A Three-Model Secure Portal

GenAI.mil's purpose is straightforward: bring the strongest commercial models into an environment that meets DoD security standards. ChatGPT Mil is accredited at Impact Level 5 (IL5), the government's cloud security bar for Controlled Unclassified Information (CUI), and supports chat, files, projects, and custom GPTs for document-heavy work like planning, policy, logistics, and administration. Grok for Government leans into reasoning, offering Auto, Fast, and Expert modes aimed at tasks from acquisition analysis to supply-chain management.

Why Is Claude Missing?

Notably absent is Anthropic's Claude — not for technical reasons, but because of an active lawsuit. In March, the DoD designated Anthropic a "supply chain risk" after the company refused to grant unrestricted access to Claude for autonomous weapons and domestic mass surveillance use cases, insisting on keeping its own safety guardrails. A federal judge ruled in late August that the designation was "illegal and baseless," finding the Pentagon's core claims about Claude "entirely unfounded." Anthropic remains excluded from DoD contracts while litigation continues. The GenAI.mil lineup, in other words, reflects not just model capability but each vendor's stance on how much safety control it's willing to cede to a government customer.

Grok's Safety Problem: Reportedly "No Reliable Fix" for CSAM

More concerning: multiple outlets report that xAI engineers internally acknowledged there is no reliable technical fix for Grok generating child sexual abuse material (CSAM) — a byproduct of the same capability that lets the model produce explicit adult content, since shifting a prompt's described subject to a minor triggers the same generation pathway. Despite this, Grok for Government received its IL5 provisional authorization and was deployed to 3 million DoD users with no public risk disclosure on the issue. xAI currently holds a $200 million DoD contract.

What This Means for Builders and Enterprises

  • Government adoption ≠ safety validated: IL5 certifies data storage and transmission security, not that content-generation risks have been resolved
  • Safety guardrails are becoming a bargaining chip: vendors willing to loosen restrictions may win large government contracts faster — that's not the same as being more mature
  • Multi-model deployment is the norm: GenAI.mil running three vendors at once shows large institutions are no longer betting on a single provider

The real story here isn't which AI to pick — it's the invisible legal and safety tug-of-war behind government AI adoption that actually decides which model gets deployed. 好不好用,試了才知道 — you only know if it works once you've tried it.

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

留言

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code