Google, Microsoft, and xAI have joined OpenAI and Anthropic in giving the US Commerce Department pre-release access to evaluate their AI models, creating voluntary oversight of all five major frontier AI labs through an office with no statutory authority and fewer than 200 staff. The expansion was catalysed by the Mythos crisis and a potential executive order that would formalise the review process.TL;DR
The Mythos crisis forced the United States government to confront a question it had been avoiding: what happens when an AI model is powerful enough to threaten national security and the government has no formal mechanism to evaluate it before the public gets access? On Tuesday, the Commerce Department announced that Google, Microsoft, and xAI have agreed to give the US government pre-release access to their AI models for evaluation. They join OpenAI and Anthropic, which have been submitting models to the same office since 2024. Five companies now account for the vast majority of frontier AI development worldwide, and all five have agreed to let a single government office test their systems before deployment. The arrangement is voluntary, has no statutory basis, and gives the government no power to block a release. It is also the closest thing the United States has to an AI oversight system, and it was built in less than two years by an office with fewer than two hundred staff.
The Center for AI Standards and Innovation sits within the Commerce Department’s National Institute of Standards and Technology. It was established under President Biden in 2023 as the AI Safety Institute, re-established under Trump with a new name and a reorientation toward standards and national security rather than safety research. The centre has completed more than 40 evaluations of AI models, including state-of-the-art systems that have never been released to the public. Developers frequently submit versions with safety guardrails stripped back so that evaluators can probe for national security-relevant capabilities: biological weapon synthesis pathways, cyberattack automation, and autonomous agent behaviours that could be difficult to control at scale.
Chris Fall now directs the centre, following the abrupt departure of Collin Burns, a former AI researcher at Anthropic who was chosen for the role but pushed out by the White House after four days. Burns had left Anthropic, given up valuable stock, and relocated across the country to take the government position. His removal, reportedly driven by his connection to a company the administration was actively fighting, illustrates the political complexity of building an oversight system for an industry where the evaluators and the evaluated come from the same talent pool. Trump’s broader AI regulatory approach has prioritised federal preemption of state regulation and a light-touch posture toward industry, but the model evaluation programme represents a harder edge: the government wants to see what these systems can do before anyone else does.
The new partnerships with Google, Microsoft, and xAI expand what had been a two-company arrangement into something closer to comprehensive frontier coverage. OpenAI and Anthropic have renegotiated their existing agreements to align with Trump’s AI Action Plan, which directs the centre to lead national security-related model assessments and positions it as part of a broader “evaluations ecosystem.” The agreements are not contracts. They are voluntary commitments that the companies can withdraw from at any time. No statute requires pre-release evaluation. No regulation gives the centre authority to delay or block deployment. The entire system depends on the AI companies deciding, for their own strategic reasons, that giving the government early access is preferable to the alternative.
The alternative, from the companies’ perspective, is legislation. Several draft bills would give the centre permanent statutory authority, mandatory evaluation requirements, and the power to impose conditions on deployment. The Pentagon has already demonstrated willingness to blacklist AI companies that refuse to comply with government demands, designating Anthropic a supply-chain risk after the company refused to allow its models to be used for autonomous weapons or mass domestic surveillance. The voluntary evaluation agreements are, in part, a way for the remaining companies to demonstrate cooperation before cooperation becomes compulsion.
The expansion of the evaluation programme is happening against the backdrop of the Mythos crisis. Anthropic’s breakthrough model, announced in April, can autonomously discover and exploit zero-day vulnerabilities in every major operating system and web browser. It has identified thousands of high-severity bugs, including vulnerabilities that existed for decades undetected. The White House has opposed Anthropic’s plan to expand access to Mythos beyond its initial consortium of launch partners. The NSA is using it despite the Pentagon’s blacklist of Anthropic. The EU is demanding access to Mythos for European cyber defence, arguing that the most consequential cybersecurity tool in existence cannot remain under the exclusive control of an American company that the American government has partially blacklisted.
Mythos demonstrated what the evaluation programme is designed to catch: a model whose capabilities have immediate national security implications that cannot be assessed after deployment. The centre’s 40-plus evaluations since 2024 presumably identified capabilities in unreleased models that informed policy decisions, but those evaluations happened under agreements with only two companies. Google’s Gemini, Microsoft’s models, and xAI’s Grok were not subject to pre-release government review until now. The new agreements close that gap, ensuring that the next model with Mythos-level capabilities, regardless of which lab produces it, reaches government evaluators before it reaches the public.
The programme’s structural weakness is obvious: it depends entirely on voluntary participation. A company that discovers its model has dangerous capabilities could, legally, decline to submit it for evaluation and release it anyway. The centre has no subpoena power, no injunctive authority, and no mechanism to compel disclosure. Its leverage is reputational and political: companies that participate signal responsibility, and companies that refuse invite regulation. But that leverage assumes the government can credibly threaten legislation, and the current administration’s stated preference for light-touch regulation weakens that threat.
Euro-area finance ministers have discussed Anthropic’s Mythos as a financial stability concern, recognising that a cybersecurity tool capable of discovering vulnerabilities in banking infrastructure has implications beyond traditional national security. The international dimension adds pressure: if the US government cannot demonstrate that it has oversight of frontier AI models developed on its soil, other governments will impose their own requirements, fragmenting the global AI market and creating compliance costs that the companies want to avoid. The voluntary evaluation programme is, in this reading, not oversight but a prophylactic against oversight: proof that the industry is cooperating, offered in exchange for continued freedom to self-govern.
The Trump administration is considering an executive order that would create a formal government review process for AI models, potentially transforming what is currently voluntary into something with regulatory teeth. A working group of tech executives and government officials would design the process, with options ranging from advisory review to mandatory pre-deployment approval. The administration’s challenge is that it simultaneously wants to accelerate AI development, maintain American competitive advantage over China, avoid burdening companies with regulation, and ensure that models with national security capabilities are subject to government review. These objectives are not fully compatible, and the voluntary evaluation programme is the current attempt to reconcile them.
AI capabilities are advancing into specialised domains at a pace that outstrips the government’s capacity to evaluate them. The centre’s 200-odd staff are assessing models that are being developed by companies with tens of thousands of researchers and hundreds of billions in capital. The asymmetry is structural: the companies will always know more about their models than the evaluators do, and the evaluation will always lag behind the frontier. What the programme provides is not comprehensive oversight. It is a window, narrow and dependent on goodwill, into what the most powerful AI systems can do before the rest of the world finds out. Five companies have agreed to keep that window open. Whether the window becomes a door, with the government able to walk through and impose conditions on what it sees, depends on whether the next Mythos-level capability arrives before or after Congress decides that voluntary cooperation is no longer enough.
Get the most important tech news in your inbox each week.