On 5 May, the US National Institute of Standards and Technology (NIST) announced that Google, Microsoft and xAI will submit their unreleased AI models for government evaluation before public deployment. With OpenAI and Anthropic already in similar arrangements, all five major frontier AI developers now let the US government test their models before release.

What CAISI will test, and why now
The testing will be done by the Center for AI Standards and Innovation (CAISI), part of the US Department of Commerce. Its evaluations cover three risk areas: cybersecurity, biosecurity and chemical weapons.
CAISI Director Chris Fall said: "Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications."
The timing follows Anthropic's Claude Mythos model, which raised serious alarm in the AI safety community weeks earlier. Its autonomous hacking capabilities pushed concern about AI and cybersecurity to a tipping point and helped prompt the White House to consider formalising the review process.
OpenAI and Anthropic set up their arrangements with CAISI nearly two years ago, under the Biden administration. The Trump administration has renegotiated those partnerships to fit the priorities of its AI Action Plan. CAISI has already carried out 40 evaluations, including of unreleased state-of-the-art models.
How the arrangement works
The agreements are voluntary. CAISI leads unclassified assessments of capabilities that may pose national security risks, and it focuses on threats that can be demonstrated rather than speculative ones. A task force at the centre also lets officials from across the federal government test models, including in classified settings, so defence, intelligence and law enforcement agencies can look at risks before a system reaches the public.
This is a middle path. Mandatory pre-market approval could slow US AI development relative to China, so the government relies on partnership and shared national security interests instead. The companies have their own reasons to agree. A major AI-enabled cyberattack or biosecurity incident could bring far more restrictive regulation, and voluntary review helps them keep credibility with policymakers while keeping their room to operate.
The European Union has taken a different route. The EU AI Act sets prescriptive compliance requirements, including documentation and third-party conformity assessments, with deadlines for high-risk systems recently pushed to 2027. The US model concentrates on pre-release security testing by government experts. The EU approach is more predictable because the rules are written down. The US approach is more flexible, but it is less clear what "passing" an evaluation means. UAE organisations with operations in both markets will have to work with both philosophies, and regional approaches such as Saudi Arabia's emerging AI governance standards add another layer.
Effects on buyers and release schedules
Three practical effects are worth tracking.
Procurement gets another question. When security teams assess a frontier model, whether it has been through CAISI review may become a factor, especially for government contractors and regulated industries. A model that has passed carries an implicit government signal on security. That is separate from any endorsement for commercial use.
Release dates may move. Labs may delay a launch to finish CAISI evaluation, particularly when a model is a large step up in capability. That could open windows in which Chinese models reach the market first.
Transparency expectations rise. CAISI results are not published, but the process itself puts pressure on labs to say more about their internal safety testing, and customers increasingly expect to know what red-teaming a model has been through.
For organisations in the UAE choosing a frontier model, I would add one question to every vendor conversation: has this model, at this version, been through pre-deployment testing by CAISI or a comparable body, and what can you tell us about the red-teaming behind it? The answer will not settle a procurement decision on its own, but a vendor that cannot answer it is telling you something.
---
Sources:
- Cybersecurity Dive: NIST will test three major tech firms' frontier AI models for cybersecurity risks
- SiliconANGLE: Google, Microsoft and xAI agree to allow government safety checks of AI models prior to release
- Washington Post: NIST will review new AI models from Google, Microsoft, xAI before release
- CNBC: Trump admin moves further into AI oversight