Three more labs, one testing arrangement
On May 5, 2026, the Center for AI Standards and Innovation, a division of NIST inside the Commerce Department, announced voluntary agreements with Google DeepMind, Microsoft and xAI. Under the arrangement CAISI will "conduct pre-deployment evaluations and targeted research to better assess frontier AI capabilities and advance the state of AI security," reviewing each company's models before they reach the public. [3]
The testing runs through an interagency task force housed at CAISI, letting officials from across government examine models in classified settings, and the effort centers on exchanging information between government and the developers rather than a one-way inspection. Microsoft's Natasha Crampton called the arrangement dependent on "close collaboration between industry and governments with deep technical and security expertise." [4]
Extending, not inventing, the arrangement
The mechanism is not new. In August 2024, NIST's US AI Safety Institute, CAISI's predecessor under the Biden administration, signed memoranda of understanding with Anthropic and OpenAI granting the institute "access to major new models from each company prior to and following their public release," with the institute providing feedback on potential safety improvements in coordination with the UK's AI Safety Institute. AISI director Elizabeth Kelly said at the time the agencies would "advance the science of AI safety." [5]
Atlas interpretation: The institute was renamed CAISI in June 2025 under the Trump administration with a more industry-focused, security-oriented mandate, but the pre-release access model it inherited from 2024 survived that change and became the template for signing up three more of the largest model developers, rather than a replacement for the earlier deals. [5][6]
National security framing, not general safety
CAISI director Chris Fall framed the expansion in national-security terms: "Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications." Reporting on the announcement tied it to the Trump administration's AI Action Plan from July 2025 and noted CAISI's own prior work evaluating a Chinese model, DeepSeek, in September 2025. [3][6]
Atlas interpretation: That framing marks a shift from 2024's language of safety research and mitigation to 2026's language of security and capability assessment, even though the underlying mechanism, pre-release access plus government evaluation, is the same one OpenAI and Anthropic had already agreed to. [5][3]
The same government stopping a rollout
Five weeks after the CAISI agreements were announced, the Commerce Department went further with two of the original 2024 signatories: a directive on June 12, 2026 suspended foreign-national access to Anthropic's Claude Fable 5 and Claude Mythos 5 and held back OpenAI's GPT-5.6 rollout on the same grounds, treating access to a frontier model as a controlled export rather than a product release. [6]
Atlas interpretation: The two actions share an agency and a rationale, national-security risk from frontier capability, but not a legal basis: the May agreements are voluntary testing arrangements the companies can decline to renew, while the June directive is a binding export restriction. The five-week gap shows the same government moving from asking to look at a model before release to deciding, on its own authority, that a model should not reach some users at all. [6]
Sources
- Trump admin moves further into AI oversight, will test Google, Microsoft and xAI models
CNBC · May 5, 2026
- Microsoft, Google and xAI will let the government test their AI models before launch
CNN · May 5, 2026
- Commerce AI center will evaluate Google Deepmind, Microsoft and xAI models
Nextgov/FCW · May 5, 2026
- NIST will test three major tech firms' frontier AI models for national security risks
Cybersecurity Dive · May 6, 2026
- U.S. AI Safety Institute Signs Agreements Regarding AI Safety Research, Testing and Evaluation With Anthropic and OpenAI
NIST · Aug 29, 2024
- US government agency to safety test frontier AI models before release
CIO · May 6, 2026