Five Frontier AI Labs Agree to Voluntary Pre‑Release Model Reviews by U.S. Government
Key Points
- Google, Microsoft, xAI, OpenAI and Anthropic now give the U.S. Commerce Department pre‑release access to their newest AI models.
- The Center for AI Standards and Innovation, part of NIST, conducts the evaluations with a staff of fewer than 200.
- Agreements are voluntary; companies can withdraw at any time and the center cannot block model releases.
- The expansion follows the Mythos crisis, where Anthropic’s model exposed thousands of zero‑day vulnerabilities.
- Critics warn the program lacks subpoena power and depends on companies’ goodwill, leaving a regulatory gap.
- International pressure, especially from the EU, could drive separate oversight regimes if U.S. oversight remains voluntary.
- The Trump administration is considering an executive order that could grant the center statutory authority.
Google, Microsoft, xAI, OpenAI and Anthropic have signed on to give the U.S. Commerce Department’s Center for AI Standards and Innovation pre‑release access to their newest models. The voluntary arrangement, built by a staff of fewer than 200, provides the closest approximation the United States has to an AI oversight system, though it carries no statutory authority and cannot block deployments. The expansion follows the so‑called Mythos crisis, which highlighted the need for early government scrutiny of powerful AI capabilities.
The U.S. Commerce Department announced Tuesday that Google, Microsoft and Elon Musk’s xAI have joined OpenAI and Anthropic in granting the agency pre‑release access to their latest artificial‑intelligence models. The move expands a voluntary evaluation program that now covers the five companies responsible for the bulk of frontier AI development worldwide.
The Center for AI Standards and Innovation, housed within the National Institute of Standards and Technology, coordinates the reviews. Established under President Biden in 2023 as the AI Safety Institute and later renamed by the Trump administration, the center has completed more than 40 assessments of cutting‑edge systems that have never been released to the public. Evaluators receive stripped‑down versions of models so they can probe for capabilities that could threaten national security, such as automated cyber‑attack planning or the synthesis of biological weapons.
Unlike formal regulatory schemes, the agreements are not contracts. Companies can withdraw at any time, and the center has no power to halt a model’s deployment. Its leverage rests on reputational pressure and the prospect of future legislation that could impose mandatory reviews. The voluntary framework emerged after the Mythos crisis, when Anthropic’s breakthrough model demonstrated the ability to discover zero‑day vulnerabilities across major operating systems and browsers, raising alarms in the White House and the Pentagon.
Anthropic’s Mythos model, unveiled in April, sparked a debate over whether the government should have a formal mechanism to evaluate AI systems before they reach the market. The center’s early assessments of Mythos‑level capabilities informed policy decisions, but only two companies—OpenAI and Anthropic—had been subject to review until now. Google’s Gemini, Microsoft’s Azure AI models, and xAI’s Grok will now undergo the same scrutiny, closing a gap that could have left the next high‑risk system unchecked.
Critics note the program’s structural weakness: participation is entirely voluntary, and the center lacks subpoena power or the ability to impose injunctions. If a lab determines its model poses no danger—or simply prefers to avoid government scrutiny—it could release the system without further oversight. The administration’s preference for light‑touch regulation means the threat of mandatory legislation remains uncertain.
International pressure adds another layer of complexity. European finance ministers have flagged the cybersecurity implications of models like Mythos for financial stability, while the EU has demanded access to such tools for its own cyber‑defense needs. Without a clear U.S. oversight mechanism, foreign regulators may impose their own requirements, potentially fragmenting the global AI market.
President Trump’s team is reportedly weighing an executive order that would formalize the review process, giving the center statutory authority and possibly the power to delay or condition deployments. The order would aim to balance rapid AI innovation, U.S. competitiveness with China, and national‑security concerns—a delicate trade‑off that the current voluntary program only partially addresses.
For now, the five labs keep a narrow window open for government evaluators. Whether that window expands into a door depends on future legislative action and the emergence of another model with capabilities comparable to Mythos.