AI模型上市前審查框架TRAINS：白宮8月1日大限逼近 | AI Pre-Release Review Framework TRAINS: Aug 1 Deadline
By Kit 小克 | AI Tool Observer | 2026-07-29
🇹🇼 AI模型上市前審查框架TRAINS：白宮8月1日大限逼近
AI模型上市前審查制度即將在美國正式成形——白宮與OpenAI、Anthropic、Google、Microsoft、xAI五大實驗室已就代號「TRAINS」的自願性審查框架接近達成共識，最快8月1日就會正式宣布。這個時間點並非巧合，正好卡在川普總統6月2日簽署的AI與資安行政命令60天大限上。這套機制名義上「自願」，但業界普遍認為一旦上路就會變成事實上的產業標準，不參加的實驗室反而會被市場與政府採購方質疑安全性不足。
TRAINS框架是什麼？AI模型上市前審查怎麼運作？
TRAINS是聯邦政府與前沿AI實驗室之間的「上市前30天審查窗口」機制，要求frontier模型在正式對外釋出前，先交由財政部、國防部與國土安全部組成的審查小組進行為期30天的測試評估。目前OpenAI、Anthropic、Google、Microsoft與xAI都已同意加入，Meta則尚未表態參與。
為什麼現在會有這個審查框架？
這套機制源自川普總統6月2日簽署的AI與資安行政命令，命令要求財政部、國防部與國土安全部在60天內，針對先進AI模型建立一套基準測試與自願框架，而8月1日正是這60天期限的到期日。
什麼是共享越獄嚴重度評分系統？
這次框架最受矚目的環節，是五大實驗室同意採用第一套「共享越獄嚴重度評分系統」，概念類似資安界行之有年的CVSS（通用漏洞評分系統），用統一標準為AI模型的越獄風險打分數，讓不同實驗室的安全等級可以互相比較，也讓監管機關有共通語言可以對話。
「自願」卻形同「強制」？TRAINS框架的爭議
媒體以「紙上自願、實務強制」形容這套框架的處境——條文上寫著自願參與，但一旦五大實驗室都簽署，任何拒絕加入審查的公司都會在政府採購資格與企業客戶信任度上處於劣勢，等於變相被迫跟進。
更值得注意的是時間點——歐盟AI法案（EU AI Act）針對高風險系統的完整規範將在8月2日正式生效，剛好卡在TRAINS框架宣布隔天。也就是說，全球AI監管在48小時內同時迎來兩套新規則，跨國經營的AI公司合規壓力將明顯增加。
常見問題 FAQ
Q：TRAINS框架是強制性法規嗎？
A：名義上是自願參與，但五大實驗室都已同意加入，實務上等同產業標準，未參與者可能面臨市場與政府採購壓力。
Q：哪些公司加入了TRAINS？
A：OpenAI、Anthropic、Google、Microsoft、xAI已同意參與，Meta目前尚未表態。
Q：TRAINS框架和EU AI Act有什麼關係？
A：兩者時間點相近但無直接關聯，EU AI Act高風險系統規範8月2日生效，恰好在TRAINS宣布隔天，形成全球監管疊加效應。
好不好用，試了才知道。
🇺🇸 AI Pre-Release Review Framework TRAINS: Aug 1 Deadline
An AI model pre-release review system is about to take formal shape in the US. The White House and five frontier AI labs — OpenAI, Anthropic, Google, Microsoft, and xAI — are nearing agreement on a voluntary review framework codenamed TRAINS, with an announcement expected as early as August 1. That date is no coincidence: it marks the 60-day deadline set by President Trump's June 2 executive order on AI and cybersecurity. The framework is voluntary on paper, but most in the industry expect it to function as a de facto standard — any lab that skips it risks looking less safe to regulators, enterprise buyers, and the federal government itself.
What Is the TRAINS Framework?
TRAINS is a 30-day pre-release review window between the federal government and frontier AI labs. Before a frontier model ships publicly, a review team drawn from Treasury, Defense, and Homeland Security gets 30 days to test and evaluate it. OpenAI, Anthropic, Google, Microsoft, and xAI have all agreed to participate; Meta has not.
Why Is This Review Framework Happening Now?
The framework traces back to Trump's June 2 executive order on AI and cybersecurity, which gave Treasury, Defense, and Homeland Security 60 days to build a benchmarking process and voluntary framework for advanced AI models. August 1 is that 60-day deadline.
What Is the Shared Jailbreak Severity Scoring System?
The most notable piece of the deal is that the five labs have agreed to adopt the first shared jailbreak severity scoring system — modeled on CVSS, the widely used vulnerability scoring standard in cybersecurity. It gives regulators and labs a common scale for how severe a jailbreak risk is, making safety levels comparable across different AI models for the first time.
Voluntary in Name, Mandatory in Practice?
Coverage of the deal has framed it as “voluntary on paper, mandatory in practice” — once all five major labs sign on, any company that opts out faces real disadvantages in government procurement and enterprise trust, effectively pressuring holdouts to join anyway.
Timing adds another layer: the EU AI Act's full requirements for high-risk AI systems take effect August 2 — one day after the TRAINS announcement is expected. That puts two major new AI regulatory regimes into force within 48 hours of each other, raising compliance pressure for any AI company operating globally.
FAQ
Q: Is TRAINS legally mandatory?
A: It's voluntary on paper, but with all five major labs signing on, it functions as an industry standard in practice.
Q: Which companies have joined TRAINS?
A: OpenAI, Anthropic, Google, Microsoft, and xAI have agreed to participate; Meta has not yet committed.
Q: How does TRAINS relate to the EU AI Act?
A: They're unrelated but land almost simultaneously — the EU AI Act's high-risk rules take effect August 2, one day after the expected TRAINS announcement.
好不好用，試了才知道 — as always, you don't know if a framework works until it's actually tested in practice.
Sources / 資料來源
- AI Model Safety Standards Deal Targets August 1: Five Labs Adopt First Jailbreak Scoring Scale
- Voluntary on Paper, Mandatory in Practice: White House AI Review Hits August 1 Deadline
- White House and Top AI Labs Near Deal on Voluntary Frontier-Model Standards
常見問題 FAQ
TRAINS框架是強制性法規嗎？
名義上是自願參與，但五大實驗室都已同意加入，實務上等同產業標準，未參與者可能面臨市場與政府採購壓力。
哪些公司加入了TRAINS？
OpenAI、Anthropic、Google、Microsoft、xAI已同意參與，Meta目前尚未表態。
TRAINS框架和EU AI Act有什麼關係？
兩者時間點相近但無直接關聯，EU AI Act高風險系統規範8月2日生效，恰好在TRAINS宣布隔天，形成全球監管疊加效應。
延伸閱讀 / Related Articles
- AI推理雲估值大戰：Fireworks 175億、Baseten 130億同月爆發 | AI Inference Cloud Valuations Explode: Fireworks $17.5B, Baseten $13B
- 開放權重AI論戰：輝達聯署20企業，Anthropic表態不封殺 | Open-Weight AI Fight: Nvidia Rallies 20 Firms vs Anthropic
- Cognition併購Poke：AI人格戰取代跑分競賽 | Cognition Acquires Poke: The AI Personality War Begins
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