Google DeepMind, Microsoft, and xAI have agreed to allow the US government to conduct pre-deployment security evaluations of their AI models. The agreements were announced on May 5, 2026, by the Commerce Department's Center for AI Standards and Innovation (CAISI), which has conducted 40 such reviews to date.
Key Takeaways
- Google DeepMind, Microsoft, and xAI signed new agreements with CAISI
- CAISI has conducted 40 AI model pre-deployment evaluations so far
- OpenAI and Anthropic renegotiated their original 2024 agreements
- The White House is considering an executive order mandating government review of new AI models
- CAISI operates under the Department of Commerce as part of NIST
What Is CAISI and What Does It Evaluate?
The Center for AI Standards and Innovation (CAISI) is a unit of the National Institute of Standards and Technology (NIST) within the US Department of Commerce. Its mission is to conduct independent scientific evaluations of AI models for national security implications and technical capabilities — before they are released to the public.
In the May 5, 2026 announcement, CAISI director Chris Fall stated: "Independent, rigorous measurement science is essential to understanding frontier AI and its national security implications. These expanded industry collaborations help us scale our work in the public interest at a critical moment." CAISI has conducted 40 reviews to date. The first agreements — with OpenAI and Anthropic — were signed in August 2024.
The new agreements with Google DeepMind, Microsoft, and xAI expand the circle of companies covered by this mechanism. Meanwhile, OpenAI and Anthropic have renegotiated their previous agreements to align with the priorities of the Trump administration's AI Action Plan, announced in early 2025.
Scope and Nature of Evaluations
The agreements cover "pre-deployment evaluations and targeted research to better assess frontier AI capabilities," according to the official CAISI press release. No details were provided about which specific models will be evaluated or how much time CAISI will have to complete a review before a model's public launch.
Notably, the agreements are voluntary — no law requires AI companies to submit to such evaluations. This is part of a broader policy stance in which the Trump administration encourages industry self-regulation rather than imposing mandatory requirements. By comparison, the EU's approach under the 2024 AI Act mandates risk assessments for "high-risk" and general-purpose AI systems with significant capabilities, conducted by developers and certified by notified bodies under the European Commission.
Executive Order Under Consideration
A May 4, 2026 New York Times report revealed that the White House is considering an executive order that would "bring together tech executives and government officials" to oversee new AI models. The option under consideration would give the government pre-release access to models without blocking their launch.
According to the NYT, some officials fear "political repercussions if a devastating AI-enabled cyberattack were to occur." The context is the recent launch of Anthropic's Mythos model — a system with significant cybersecurity capabilities that raised concerns in government circles.
The CAISI mechanism is seen as a path of least resistance: voluntary but visible industry-government cooperation. Agreements with Google DeepMind, Microsoft, and xAI may be an attempt to get ahead of stricter regulations.
Why This Matters
The expansion of CAISI agreements to include Google DeepMind, Microsoft, and xAI signals that leading AI companies are willing to accept government pre-release oversight — even in a pro-industry political environment. This reverses the 2024–2025 trend of the White House actively rolling back AI safety regulations.
From a geopolitical perspective, the agreements carry contextual significance: the US and China are engaged in intensive competition in frontier AI. The ability for a government body to evaluate new model capabilities — particularly in cybersecurity — functions as a de facto technical intelligence mechanism. For AI companies, accepting such evaluations means balancing government relationships against investor expectations for rapid product releases.
The key question remains whether the voluntary nature of these agreements will hold if the Trump administration decides to issue an executive order applying to all companies. In that scenario, firms that signed with CAISI early will have a stronger negotiating and operational position.
What's Next?
- The White House has yet to set a timeline for a potential executive order — per NYT, it is being considered in the context of growing cybersecurity concerns
- CAISI did not disclose which specific models are covered by the new agreements or the review timelines
- OpenAI and Anthropic have operated under CAISI since 2024 — the new agreements with Google DeepMind, Microsoft, and xAI mean five leading US AI labs are now covered by this mechanism
Sources
- The Verge — Google, Microsoft, and xAI will allow the US government to review their new AI models
- NIST / CAISI — CAISI signs agreements regarding frontier AI national security testing
- Bloomberg — AI firms agree to give US early access to evaluate their models
- The New York Times — White House considers vetting AI models before they are released
