Microsoft, Google, xAI Will Give US Early Access to AI Models

Microsoft, Google, and xAI announced Tuesday they will give the U.S. government pre-release access to new artificial intelligence models, allowing federal evaluators to test for national security risks before the models reach the public.

The agreements run through the Center for AI Standards and Innovation (CAISI), a unit inside the U.S. Department of Commerce. CAISI said the partnerships will let it evaluate frontier models before deployment and run research on capabilities and security risks.

This is the first time three of the largest American AI developers have committed to pre-release government testing in a single coordinated move - and it brings the policy framework largely in line with the existing arrangements that OpenAI and Anthropic agreed to in 2024.

What Was Announced

CAISI now has formal pre-release access agreements with five major AI companies: Microsoft, Google, xAI, OpenAI, and Anthropic. The new memoranda of understanding signed Tuesday by Microsoft, Google, and xAI extend an approach that was previously limited to two labs.

Under the agreements, the companies will share unreleased models with CAISI evaluators in advance of public launch. The center can probe for capabilities relevant to national security - things like cyberattack assistance, chemical and biological knowledge, and large-scale autonomous behavior - before those capabilities ship to customers. The companies retain the right to launch on their own timelines, but the government gets a structured look first.

OpenAI and Anthropic also used Tuesday's announcement to renegotiate their existing partnerships with CAISI. The updated framework aligns the older agreements with the priorities laid out in President Donald Trump's AI Action Plan, the administration's flagship policy document on artificial intelligence.

Why This Matters

Pre-release evaluation has been one of the most contested questions in AI policy for the past two years. Advocates argue that frontier models can cross capability thresholds with little warning, and that some form of independent testing is the only practical way to catch dangerous emergent behavior before public deployment. Critics, including some developers, have worried that government access could slow shipping cadence or leak proprietary research.

Tuesday's deal threads that needle. The companies are not required to delay launches, and the agreements are voluntary rather than statutory. But by formalizing the arrangement across five major labs at once, the administration creates de facto industry norms - any lab that opts out will look like an outlier.

This is also a notable continuity moment. The CAISI program, originally launched as the AI Safety Institute under the Biden administration, was widely expected to be downsized or dismantled under Trump. Instead, it has been renamed and refocused around the AI Action Plan, and is now expanding its reach.

Who's Affected

The most direct impact is on the developers themselves. Microsoft, Google, and xAI all run aggressive release cadences across their respective foundation models - GPT-class models inside Azure, Gemini's frontier line, and the Grok family. Each will now need to bake CAISI evaluation into release planning. In practice, this likely means a 1-3 week pre-release evaluation window, similar to what OpenAI and Anthropic have already been operating under.

Enterprise customers benefit indirectly. The presence of independent capability evaluation gives risk-averse buyers - banks, healthcare systems, government agencies - more confidence that frontier models have been stress-tested for the most concerning failure modes before they show up in production deployments.

For competitors outside this group of five, the agreements raise a strategic question. Does opting out of CAISI evaluation become a competitive disadvantage when selling into regulated industries or government? It is too early to say, but the pattern is clear.

Industry Reaction

Industry observers framed Tuesday's announcement as a pragmatic compromise. Several noted that the Trump AI Action Plan had emphasized American AI leadership and avoiding heavy-handed regulation - and that pre-release evaluation, structured as a voluntary security partnership rather than a regulatory requirement, fits cleanly within that frame.

The companies themselves struck similar notes. The agreements are described as enabling national security review rather than as licensing or approval regimes. That framing matters: it lets the labs preserve commercial control over launch decisions while still giving the government a meaningful look at what's coming.

There is, however, an open question about scope. CAISI's evaluation capacity is finite. With five major labs now feeding the center pre-release models on rolling cadences, the practical bottleneck may be how quickly the government can actually conduct these reviews.

What's Next

Several things to watch in the coming months. First, the published evaluation methodology. CAISI has previously released summaries of its testing approaches, and similar documentation for the new agreements would clarify what "national security review" actually entails in practice.

Second, whether other developers - Meta, Mistral, Cohere, and the major Chinese labs - sign on or stay out. Meta's Llama models are widely deployed inside U.S. enterprises and would be a particularly notable addition.

Third, the international dimension. The U.K. AI Safety Institute and similar bodies in the EU and Japan have been negotiating their own pre-release access agreements. Tuesday's announcement could accelerate parallel deals abroad, or it could reinforce a U.S.-first approach that other governments push back on.

Bottom Line

This is the clearest signal yet that pre-release government evaluation is becoming a standard feature of the U.S. AI industry, not a niche commitment by a few labs. For developers, the practical takeaway is that release planning now needs to account for federal review windows. For enterprises and end users, the upshot is that the most powerful models being shipped in 2026 are increasingly going through a security check before they reach the wild.