2026年6月12日，下午五點二十一分
美國商務部長 Lutnick 在美東時間 17:21 簽署出口管制指令，要求 Anthropic 立即停止所有外國國籍人員存取 Fable 5 與 Mythos 5。AWS 同步切斷外國用戶的連線。台灣開發者的 API 請求開始回傳錯誤，但那一刻多數人還不知道發生了什麼事。
這是美國史上首次以出口管制工具直接鎖定民用 AI 模型。以往管制的是晶片、是硬體、是製程設備——這次管的是模型本身，是跑在雲端、靠 token 計費、任何人持信用卡就能叫用的推論服務。
「AI 模型等同戰略科技出口」這個先例，就這樣成立了。
觸發點與後來補充的細節
官方理由是國家安全：Mythos 5 在紅隊測試中突破幾乎所有 NSA 機密系統。Anthropic 反駁，稱那個越獄技術並不特殊，GPT-5.5 同樣具備相同能力，不構成特殊國安威脅。
後來披露的一條線索更耐人尋味：據 MIT Technology Review 的分析，Amazon CEO Andy Jassy 曾致電財政部長 Bessent，指控 Fable 5 構成國安威脅——而 Amazon 既是 Anthropic 的最大股東，同時正在自建競爭 AI 模型。出口管制作為企業博弈工具的可能，在這裡浮現得格外清晰。
從觸發到部分解封，過了兩週。6月27日，Lutnick 與財政部長 Bessent 出面斡旋，Mythos 5 在 Project Glasswing 框架下恢復限定部署，僅限網路安全研究，適用對象為百餘家美國「可信夥伴」機構。Fable 5 談判仍在進行。同期，OpenAI GPT-5.6 也以「分批發布」模式上線，Trump 政府要求先限定可信美國夥伴存取，OpenAI 表達異議但配合了。
新範式正在成形：高能力模型→政府審查→「可信夥伴」優先開放→逐步擴大。這不是臨時應急，這是制度。
台灣在哪一格
台灣用戶在這次停服範圍之內，這是已確認的事實。至於台灣是否進入「可信夥伴」名單，目前沒有公開答案。
問題的結構比「進不進名單」更深：一旦頂級模型的存取權變成政府逐案審核的特許，台灣的 AI 研究機構、新創公司、以 Claude API 為核心的產品，都得在不確定裡重新評估技術棧。依賴 Anthropic 最高階模型的程度越深，風險敞口越大。
前白宮 AI 顧問 Dean Ball 說得直白：「美國政府任意撤除最先進模型在所有外國的使用權，表明 AI 出口策略對決策者已無參考價值。」這句話的意思是，策略規劃的假設基礎——「只要合規就能存取」——已經不成立。
開源那條路跑得多快
每次出口管制，開源就往前走一步。DeepSeek R1 在美國晶片管制後以 MIT 授權開源，效能對標 ChatGPT-o1；DeepSeek V4 在 H200 進一步封鎖後開源，SWE-bench Verified 達 80.6%，推論成本約為 GPT-5.5 的七分之一。GLM-5.2 同樣可下載自建，無任何管控。
這條路線的邏輯很簡單：出口管制鎖的是服務端點，鎖不住已經流通的模型權重。Zoho CEO Sridhar Vembu 在這波事件中直接宣告「全球化已死」，建議企業轉向開源 AI。建議主要針對印度，但這個判斷對台灣同樣適用。
問題是，開源替代方案在特定任務上確實縮短了差距，但前沿模型的能力邊界仍在快速移動。選擇開源路線是降低地緣依賴，但不是免費的——維運成本、微調需求、安全紅隊，這些都得自己扛。
下次不會有兩週緩衝
這次從停服到部分解封花了兩週，但那是因為 Anthropic 有足夠的政治資本和法律資源去打這場仗。下一個被列管的模型，或者下一個被列管的公司，不一定有同樣的條件。
更需要盯的問題：這個框架會不會擴展到 OpenAI 或 Google 的高端模型？GPT-5.6 的「分批發布」機制已經在相同的政府壓力下成形。答案大概不是「會不會」，而是「什麼時候、哪個模型、哪個觸發條件」。
對台灣開發者來說，現在需要知道的具體問題只有一個：手上產品的關鍵推論路徑，有多少比例跑在無法自控的美國前沿模型 API 上？
— 邱柏宇
延伸閱讀
API Keys Became Strategic Materiel Overnight
June 12, 2026, 5:21 PM Eastern
Commerce Secretary Lutnick signed the order. Anthropic immediately suspended access to Fable 5 and Mythos 5 for all foreign nationals — including its own non-American employees. AWS cut connections to foreign users in parallel. Somewhere in Taiwan, API calls started returning errors.
This was the first time the United States had used export control authority to directly restrict a commercial AI model. Not chips. Not fabrication equipment. The model itself — the inference endpoint, billed per token, accessible to anyone with a credit card — became a controlled strategic export.
That precedent is now set.
The Trigger, and What Came After
The official justification: Mythos 5 breached nearly every NSA classified system in red-team testing. Anthropic pushed back, arguing the jailbreak technique was unremarkable and that GPT-5.5 had equivalent capabilities. The government did not revise its position.
A more complicated picture emerged in reporting from MIT Technology Review: Amazon CEO Andy Jassy had called Treasury Secretary Bessent to flag Fable 5 as a national security threat. Amazon is Anthropic’s largest shareholder. Amazon is also building competing AI models. Export controls as a competitive instrument is not a conspiracy theory at this point — it’s a plausible reading of the documented sequence.
Two weeks later, on June 27, Lutnick and Bessent brokered a partial resolution. Mythos 5 resumed limited deployment under the Project Glasswing framework, restricted to cybersecurity research at over 100 approved U.S. “trusted partner” institutions. Fable 5 negotiations remained ongoing. Simultaneously, OpenAI’s GPT-5.6 launched on June 26 under a “staged release” model — trusted American partners first, broader access later. OpenAI objected to the terms. It complied anyway.
The pattern is now a policy: high-capability model → government review → trusted-partner gating → gradual expansion. This is not emergency management. This is a regulatory architecture.
Where Taiwan Stands
Taiwan users were explicitly within the scope of the suspension. Whether Taiwan will appear on any “trusted partner” list has no public answer yet.
The structural problem runs deeper than list membership. Once access to frontier models becomes a government-issued permit reviewed case by case, every Taiwanese AI research institution and startup that built on Claude’s top-tier API is operating on uncertain ground. The dependency on Anthropic’s highest-tier models is the exposure.
Former White House AI advisor Dean Ball framed it plainly: “The U.S. government arbitrarily removing the most advanced models from all foreign access shows that AI export strategy has no reference value for decision-makers.” The planning assumption — that compliance equals access — no longer holds.
How Fast Open Source Moves
Every export control cycle, open-source moves forward. DeepSeek R1 launched under MIT license after U.S. chip controls tightened, matching ChatGPT-o1 on benchmarks through architectural efficiency rather than raw compute. DeepSeek V4 followed after H200 restrictions expanded, reaching 80.6% on SWE-bench Verified at roughly one-seventh the inference cost of GPT-5.5. GLM-5.2 is downloadable and self-hostable with no access controls attached.
The logic is structural: export controls lock the service endpoint, not model weights already in circulation. Zoho CEO Sridhar Vembu declared “globalization is dead” during this episode and advised enterprises to shift toward open-source AI. The audience was primarily Indian firms, but the calculus applies equally to Taiwan.
Open-source alternatives close the gap on specific tasks. Frontier model capabilities, however, keep moving. Choosing the open-source path reduces geopolitical dependency — it doesn’t eliminate operational cost. Fine-tuning, red-teaming, and inference infrastructure all transfer to the team running the model.
The Next One Won’t Come With Two Weeks’ Warning
The two-week gap between the suspension and partial restoration existed because Anthropic had the legal resources and political capital to contest the order. The next company on the receiving end may not. The next model — potentially from OpenAI or Google’s high-end tier — may not get the same negotiating window. GPT-5.6’s staged-release structure already formed under the same government pressure that triggered the Anthropic order.
The specific question worth answering now: what fraction of a product’s critical inference paths run through U.S. frontier model APIs that cannot be self-hosted or replaced within 48 hours?
That number is the actual risk exposure. Everything else is commentary.
— 邱柏宇
Related Posts
https://justfly.idv.tw/s/nk27Puw