The White House is asking OpenAI to slow roll the release of its new model over safety concerns

The Trump administration has asked OpenAI to limit the initial release of its newest model, GPT 5.6, to a select group of close partners rather than making it publicly available — marking the most direct government intervention in AI deployment since the technology exploded into the mainstream in late 2022.

According to The Information, CEO Sam Altman told staff this week that the government would be "approving access customer by customer" during a preview period. If the limited release goes well, OpenAI hopes to follow with a broader public release a few weeks later.

The agencies behind the request — the Office of the National Cyber Director and the Office of Science and Technology Policy — reportedly worked closely with OpenAI on the upcoming release. The intervention comes just weeks after Trump signed an executive order directing certain AI companies to voluntarily submit new models to the government for testing and evaluation before releasing them publicly.

The move represents a striking reversal for an administration that originally positioned itself as taking a hands-off approach to AI regulation. It also puts OpenAI in unfamiliar territory: for the first time, the company's product release timeline is being shaped not by competitive dynamics or technical readiness, but by government approval.

**The Anthropic precedent**

The administration's request to OpenAI mirrors a decision Anthropic made voluntarily earlier this year. Anthropic's frontier cybersecurity model, Claude Mythos, was released only through a restricted program called Project Glasswing, limited to a small group of vetted partners. Anthropic argued the model was too powerful for unrestricted public access, citing its ability to identify and exploit software vulnerabilities at speeds no human could match.

The question of whether Anthropic's caution was genuine safety concern or sophisticated marketing has been debated ever since. But the practical effect — keeping a powerful model out of public hands — is now being replicated by the government with OpenAI, without the same voluntary framing.

The difference matters. When a company decides to restrict its own release, it is exercising commercial judgment. When the government directs a company to restrict its release, it is exercising regulatory authority — even if that authority is currently informal and relies on cooperation rather than compulsion.

**The cybersecurity argument**

The specific concern driving the intervention is cybersecurity. Large language models have demonstrated an ability to write functional malware, identify software vulnerabilities, and in some cases execute attacks autonomously. Frontier models with advanced coding capabilities could theoretically discover and exploit zero-day vulnerabilities — previously unknown security flaws — faster than defenders can patch them.

This is not hypothetical. Security researchers have documented LLM-assisted cyberattacks increasing in sophistication over the past two years. The models themselves are not yet autonomous threat actors, but they function as force multipliers for both attackers and defenders, and the offense-defense balance in cybersecurity has historically favored attackers who need only find a single vulnerability while defenders must protect every entry point.

For a model like GPT 5.6, which presumably represents a significant capability leap over its predecessors, the government's concern is that public release could provide malicious actors with a tool that dramatically lowers the cost and increases the speed of sophisticated cyberattacks.

**The free speech counterargument**

The government's intervention raises immediate First Amendment concerns. Code is speech — this was established in the 1990s encryption export debates, when the government attempted to restrict the publication of encryption source code and was overruled by courts on free speech grounds.

If OpenAI's model can generate code, and the government is restricting who can access that model, the question becomes whether the government is regulating the distribution of speech. The legal framework for this is unsettled. The administration is not banning the model's release; it is requesting a delayed, limited release. But the distinction between a request and a directive becomes blurry when the requesting party is the federal government and the recipient depends on government contracts, regulatory goodwill, and future policy decisions.

OpenAI has little incentive to resist. The company is preparing for an IPO, navigating an increasingly complex regulatory environment, and competing with Anthropic for both government contracts and consumer market share. Cooperation with the administration's request costs OpenAI a few weeks of broad release and gains it regulatory goodwill that could prove valuable in future policy negotiations.

**The global competitiveness question**

There is a broader strategic dimension that the administration's intervention does not address. OpenAI and Anthropic are American companies, but the frontier of AI development is global. Chinese labs are advancing rapidly. European regulators are developing their own frameworks. If the United States government begins restricting the release of its most capable models, it does not eliminate the technology from the world — it merely ensures that the most powerful AI tools available to the public are developed and distributed by companies outside American jurisdiction.

The administration's approach implicitly assumes that American companies will comply with government requests and that this compliance will enhance national security. But a malicious actor seeking a powerful AI model will not limit their search to OpenAI. They will find alternatives — many of which are already available as open-source models with no government oversight whatsoever.

**What this means for you**

If you are a developer or business that relies on access to cutting-edge AI models, the era of immediate public availability may be ending. GPT 5.6's restricted preview could become the template for future releases, where government approval determines which customers get access first and which wait. Plan for longer evaluation periods and consider diversifying your model dependencies across multiple providers.

If you work in cybersecurity, the government's intervention validates what you already know: AI is reshaping the threat landscape. Whether this particular model is restricted or not, the capability trajectory is clear. Invest in AI-assisted defensive tools and threat detection systems that match the speed and sophistication of AI-assisted attacks.

If you follow AI policy, this is the moment where voluntary self-regulation gave way to government involvement in model deployment. The administration that promised hands-off oversight is now directly influencing release timelines. The question is no longer whether the government will regulate frontier AI — it is how, how broadly, and whether the framework will be legislative or ad hoc. Watch for the executive order's implementation details and for any congressional action that codifies these restrictions into law. Ad hoc requests to individual companies are not a sustainable governance model, but they are the model we have for now.