五角大廈AI評測：Anthropic拒簽自主武器條款遭封殺 | Pentagon AI Review: Anthropic Blocked Over Weapons Clause

9月 08, 2026

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

🇹🇼 五角大廈AI評測：Anthropic拒簽自主武器條款遭封殺

五角大廈AI合約最近因為一份FOIA訴訟文件曝光而成為焦點：OpenAI、Anthropic、Google、xAI四家公司自2025年7月起，各自與美國國防部簽下上限2億美元的合約，協助打造軍用AI。但其中Anthropic的合約在今年談崩，原因很直接——五角大廈要的自由度，Anthropic不肯給。

合約卡在哪：「任何合法用途」條款

今年稍早，國防部要求Anthropic把Claude部署到內部AI平台GenAI.mil，並簽署一份允許軍方「任何合法用途」使用其模型的後續合約。Anthropic執行長Dario Amodei當場反對，堅持合約必須明確排除全自主武器系統與境內大規模監控兩項用途。談判因此卡死，Anthropic拒絕簽字。

結果：被列入供應鏈風險黑名單

國防部長Pete Hegseth三月直接把Anthropic列為「供應鏈風險」，等於封殺其服務在政府端的使用。Anthropic告上法院，聯邦法官上個月裁定撤銷這項封殺，但傷害已經造成——五角大廈轉頭跟Google、OpenAI、xAI簽下後續協議，把AI工具部署到機密軍事網路上，Anthropic則被晾在一旁。

OpenAI那邊更露骨：軍方要「幾乎不拒答」的AI

同一批FOIA文件也揭露，OpenAI今年2月27日簽下的最新版五角大廈合約，允許其服務用於機密軍網。軍方明確要求的規格之一，是拒答率要低——白話說就是希望AI盡量不要說「不」。這跟一般企業客戶要的「安全護欄」方向完全相反，也是這次報導裡最讓人在意的細節。

對企業用戶的實際啟示

評估AI供應商做政府或國防相關案子時，得先讀清楚對方的可接受使用政策（AUP），各家在武器與監控用途上的紅線差異很大
「拒答率」不是越低越好，要看場景——軍規AI跟企業客服AI是完全不同的產品邏輯
供應商的倫理立場會直接影響你能不能用它做特定案子，選型前先確認

這場拉鋸還沒完——Anthropic贏了官司，但沒拿到合約；其他三家拿到合約，但也被攤在陽光下檢視軍規AI的具體規格。好不好用，試了才知道。

🇺🇸 Pentagon AI Review: Anthropic Blocked Over Weapons Clause

The Pentagon AI contracts story broke wide open this week after documents from a Freedom of Information Act lawsuit by The Intercept revealed that OpenAI, Anthropic, Google, and xAI each signed Department of Defense deals worth up to $200 million starting July 2025. Anthropic's follow-on deal, though, fell apart — and the reason is worth understanding if you sell software to government customers.

The Sticking Point: an "Any Lawful Use" Clause

Earlier this year, the DOD asked Anthropic to deploy Claude on its internal GenAI.mil platform under a follow-up agreement granting the military unrestricted access to the model for "any lawful use." CEO Dario Amodei pushed back, insisting the contract explicitly exclude fully autonomous weapons systems and domestic mass surveillance. Anthropic refused to sign without those carve-outs, and talks stalled.

The Fallout: a Supply Chain Risk Blacklist

In March, Defense Secretary Pete Hegseth designated Anthropic a "supply chain risk," effectively banning its services from government use. A federal judge overturned that ban last month, but the damage was already done — the Pentagon had moved on, striking follow-up agreements with Google, OpenAI, and xAI to deploy their tools on classified military networks instead.

OpenAI's Deal Is More Blunt: an AI That Rarely Says No

The same FOIA documents show OpenAI signed the latest version of its Pentagon agreement on February 27, permitting deployment across classified networks. One requirement stands out: the military explicitly wanted a minimal refusal rate — an AI tuned to say "no" as little as possible. That's the exact opposite of what most enterprise buyers ask for, and it's the detail worth paying attention to here.

What This Means If You Buy AI for Regulated Work

Read a vendor's Acceptable Use Policy before assuming it fits a government or defense contract — the red lines on weapons and surveillance differ a lot between providers
A "low refusal rate" isn't universally good or bad — it depends entirely on the use case, and military specs are not enterprise specs
A vendor's stated ethics stance can directly gate whether you can use it for certain work, so check before you commit to a stack

Anthropic won its lawsuit but lost the contract. The other three vendors kept their contracts but are now having the specifics of "military-grade AI" picked apart in public. 好不好用，試了才知道 — you only find out if it works by actually trying 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

搜尋此網誌

AI小貼士