Before the next powerful AI model reaches the public, the federal government may demand proof that it won’t help someone build a bioweapon, crack critical infrastructure, or overwhelm national security defenses. The White House is actively studying an executive order that would require developers of the most advanced artificial intelligence systems to demonstrate safety before release, according to the policy groundwork now visible across multiple federal agencies. The move would mark a striking reversal for an administration that began by dismantling its predecessor’s AI oversight framework.
The clearest signal sits inside the Commerce Department. In May 2026, the Center for AI Safety and Innovation, known as CAISI, publicly disclosed agreements with Google DeepMind, Microsoft, and xAI to conduct national security testing on frontier AI models. Under those agreements, CAISI can evaluate systems both before and after public release. The institute has already completed more than 40 evaluations, including assessments of models that have never left the companies that built them.
That track record is significant. It means the federal government is not proposing something theoretical. CAISI already has hands-on experience running the exact kind of pre-deployment review that a new executive order would formalize and make mandatory.
The infrastructure that survived a political reversal
In January 2025, President Trump formally rescinded Executive Order 14110, the Biden-era directive that had tasked the National Institute of Standards and Technology with developing AI safety standards and had contemplated using national security authorities to oversee the most advanced systems. The White House framed the move as clearing regulatory obstacles to American competitiveness and ordered the creation of a new AI Action Plan led by senior officials.
But the technical machinery that Biden’s order helped set in motion did not disappear. NIST’s AI Risk Management Framework, which defines how to map, measure, manage, and monitor AI risks across a system’s lifecycle, remained in place and continued to serve as a reference point for both federal agencies and private companies. NIST and the U.S. AI Safety Institute also circulated updated guidance on misuse risks for dual-use foundation models, the class of systems capable of being repurposed for cyberattacks, biological threats, or other dangerous applications.
Together, these documents provide a ready-made vocabulary for any future mandate. If the government orders companies to “prove” their models are safe, the definitions of what safety means, and how to measure it, already exist on paper and in practice.
What a pre-release safety requirement could look like
No draft executive order has been published, and no senior administration official has described specific requirements on the record. But the existing CAISI testing program and NIST frameworks offer a practical sketch of what compliance might involve.
Developers of frontier models would likely need to submit systems for structured evaluation before any public deployment. That process would probably include red-teaming exercises designed to probe whether a model can be manipulated into assisting with weapons development, generating sophisticated cyberattack code, or producing other outputs that cross national security thresholds. Companies would also need to provide documentation of internal risk assessments and mitigation plans showing how they intend to prevent misuse after release.
The current CAISI agreements with Google DeepMind, Microsoft, and xAI function as a voluntary pilot for this kind of regime. The open question is whether a new executive order would convert voluntary cooperation into a legal obligation, and what would happen to companies that refuse or fail the evaluation.
Notable absences and open questions
The list of companies that have signed CAISI testing agreements is telling for who is on it and who is not. OpenAI, the maker of ChatGPT and one of the most prominent frontier AI developers, is not among the disclosed partners. Neither is Anthropic, which has built its public identity around AI safety research. Whether those companies are in separate negotiations, have declined to participate, or simply have not been approached is unknown.
The internal White House process for studying the order remains opaque. There is no Federal Register notice, no confirmed list of agencies that would share enforcement authority, and no public timeline. Executive orders on complex technical subjects routinely go through months of interagency review, and some never reach the signing stage. The gap between “studying” a policy and issuing one can be enormous.
There is also an unresolved tension at the center of the administration’s approach. The same fact sheet that announced the rescission of Biden’s order emphasized removing obstacles that could slow American AI innovation. A pre-release safety mandate would, by definition, add friction to deployment timelines. Companies could face longer waits between finishing a model and launching it commercially, even as policymakers argue the United States must outpace China and other global competitors. The administration has not publicly explained how it would reconcile those competing priorities.
The state and international picture
Any federal pre-release requirement would land in a crowded regulatory environment. The European Union’s AI Act, which began phased enforcement in 2025, already imposes risk-based obligations on high-impact AI systems, including transparency requirements and conformity assessments for general-purpose models with systemic risk. The United Kingdom’s AI Safety Institute has conducted its own evaluations of frontier models. Several U.S. states are exploring or have passed AI-related legislation that could overlap with or contradict a federal standard.
A White House order that leans on CAISI and NIST could set a federal floor that simplifies compliance for companies navigating a patchwork of rules. Or it could create new conflicts if federal definitions and thresholds diverge from what states and foreign governments have established. Without published draft language, it is impossible to know which outcome the administration is aiming for.
Where this stands in June 2026
The federal government is already testing unreleased AI models against national security benchmarks. It has the institutional infrastructure, the technical frameworks, and agreements with three major developers to support a formal pre-release safety regime. What it does not yet have, at least publicly, is the binding legal instrument to make that regime mandatory.
The distance between the current voluntary system and an enforceable executive order is real but may be shorter than it appears. CAISI’s testing work has created facts on the ground that make a mandate operationally feasible in a way it would not have been even a year ago. The policy question is no longer whether the government can vet frontier AI models before release. It is whether the White House will decide it must.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
