LectureRechercheRadarCadre d'investissement
Connexion / Inscription
← Retour à la lecture
ActualitésSujet · 美国AI安全测试透明度Matériel2 min de lectureRédaction DisconfirmAI

US Government Refuses to Release AI Model Safety Testing Framework and 'Trusted Partner' List

The White House has withheld details of its voluntary safety testing rules for frontier AI models and the list of companies granted early access, citing national security.

2026-11-02 21:16 GMT+8

According to a report by German tech outlet heise online, the US government has refused to fully disclose the rules governing its voluntary safety testing of high-performance artificial intelligence (AI) models. This decision follows a Freedom of Information Act (FOIA) lawsuit filed by the civil rights organization Protect Democracy. While the Office of the National Cyber Director (ONCD) at the White House submitted some documents in response, key materials—including the full draft of the testing framework, accompanying documents, and the complete list of "Trusted Partners"—remain classified due to national security interests and protection of confidential business information.

The dispute stems from Executive Order 14409, signed by President Donald Trump in June 2026. The order established a voluntary mechanism allowing developers like Anthropic and OpenAI to grant the government access to their frontier models up to 30 days before planned release. Although the executive order explicitly states that this does not create a mandatory licensing or pre-approval requirement, the specific criteria for implementation have not been made public.

Citing documents obtained through the lawsuit, The Information reported that the actual framework likely consists of a seven-page document whose full content remains unpublished. The released materials primarily consist of emails from the ONCD and the Department of Commerce to AI labs, revealing which organizations were designated as "Trusted Partners" with early access, but leaving the selection standards and the full partner list undisclosed.

Connexion / Inscription
LectureRechercheRadarCadre d'investissement

This lack of transparency raises concerns about industry fairness and the validity of safety assessments. With the benchmarking process kept secret, external research institutions cannot independently reproduce test results, and smaller AI providers struggle to align their security measures with publicly known standards. Furthermore, "Trusted Partners" with early access may enjoy asymmetric knowledge and market advantages.

Vivian Dong, Director at Legal Advocates for Safe Science and Technology, stated that it is unacceptable for the federal government to ask the public to trust a review framework whose contents are known only to the government and select AI companies. Her organization intends to continue fighting for transparency in court. The White House did not respond to requests for comment.

Sources

  • heise.deheise online↗
Sujet

美国AI安全测试透明度

美国白宫以国家安全为由,拒绝公开针对前沿AI模型自愿性安全测试框架的具体规则及获准提前访问模型的“可信合作伙伴”企业名单。这一决定源于特朗普总统2026年6月签署的第14409号行政令,该行政令允许开发者在发布前最多30天向政府提供模型访问…

7 étapes1 piècesMis à jour 2026-11-03 03:03

Comment on en est arrivé là

  1. 2023-10-30拜登总统签署第14110号行政令,推动AI安全、可靠和可信发展,并设立NIST下属的美国AI安全研究所(AISI)。ntia.gov ↗
  2. 2024-08-01美国AI安全研究所(AISI)与Anthropic和OpenAI签署谅解备忘录,获得新前沿模型部署前的访问权限,建立协作评估框架。labs.cloudsecurityalliance.org ↗
  3. 2025-06-01特朗普政府将AI安全研究所重组为AI标准与创新中心(CAISI),重点转向网络安全、生物安全等国家安全风险。labs.cloudsecurityalliance.org ↗

Ce qui reste non résolu

  • 法院是否会在后续诉讼中强制白宫披露完整的七页安全测试框架文档?
  • 哪些具体企业被列入“可信合作伙伴”名单并获得模型提前访问权?
  • CAISI当前的评估标准如何界定“涵盖的前沿模型”,以及这些标准是否会因保密而阻碍行业合规?
Lire le sujet complet →
Continuer la recherche →