In early May, the White House blocked Anthropic from expanding access to Claude Mythos (its most capable model) from fifty organizations to one hundred and twenty. Two reasons were cited: national security, and concern that wider distribution would strain the government’s own compute allocation. Meanwhile, OpenAI is shipping GPT-5.5 to nine hundred million weekly users, and the UK’s AI Safety Institute confirmed that both models can independently complete a full multi-step cyber attack simulation from start to finish.
The capabilities that one company restricts from release because of safety concerns are now generally available through a competitor’s product. The government that blocked one company’s expansion has no mechanism to constrain the other’s. And the organizations that were denied access (universities, research labs, defense contractors) must now decide whether to use the less-restricted alternative or wait for permission that may never come.
This is not a story about AI safety. It is a story about AI governance. And the distinction matters.
The Soft License
What the White House created, without legislation or public debate, is a soft licensing regime for frontier AI capabilities. Certain models are treated as controlled technology, not through export law or classification, but through informal pressure on the companies that build them. The government decides who gets access. The companies decide whether to comply. There is no statute. There is no appeals process. There is no public record of the criteria used.
Dean Ball, a researcher who has written extensively on AI governance, observed that this arrangement cannot hold. It works only as long as one company has a capability others lack. The moment a second company reaches the same threshold (which GPT-5.5 appears to have done), the leverage evaporates. You cannot restrict access to a capability that multiple vendors offer. The soft license expires the day the moat disappears.
David Sacks, a venture capitalist and Trump administration adviser, pushed back differently. His argument: Claude Mythos does not create vulnerabilities. It discovers existing ones. The capabilities should be demystified, not mystified. He expects every leading model, including Chinese ones, to reach comparable capability within six months.
Both critiques are correct. The soft licensing regime is both untenable and misframed. But the alternative they imply (unrestricted distribution) is not obviously better. What is missing is not a stronger lock or a more open door. What is missing is the room where the decision gets made.
The Room That Does Not Exist
There is no institution in the United States that can answer the question: who should have access to frontier AI capabilities, and under what conditions?
Congress has not passed AI governance legislation. The executive branch acts through informal pressure and executive orders that change with each administration. The judiciary has no framework for evaluating AI capability claims. State legislatures (as we have documented in these pages) are building governance capacity from the ground up, but their jurisdiction does not extend to the national security dimensions of frontier models.
The result is governance by accident. The White House blocks Anthropic because it can. OpenAI ships because it can. States regulate what is within their reach. And the most consequential decisions about the most powerful technology in human history are made in rooms that do not technically exist, by people accountable to no electorate, under criteria that are not public.
The Foundation Question
This publication covers Oklahoma, energy, and AI policy. We do not pretend to have jurisdiction over the White House’s relationship with Anthropic. But we notice the pattern.
The pattern is centralization without accountability. The most capable AI systems are controlled by a small number of private companies. Access decisions are made by a small number of government officials. The public (including state governments, universities, school districts, small businesses, and civic organizations) participates in none of these decisions and learns about them after the fact.
We have written before about Foundation: a civic framework built on the premise that the AI transition requires new governance infrastructure, not just new technology. Foundation’s answer to the question of who controls the thinking machines is not “the government” or “the companies” or “the market.” It is: the institutions we have not yet built.
A Guardian AI (a public AI system operating under democratic oversight, serving citizens rather than shareholders or security agencies) does not exist. The infrastructure for one does not exist. The legislation authorizing one does not exist. But the need for one grows with every quiet decision made in rooms that do not technically exist.
The White House may have been right to block the Mythos expansion. Or it may have been wrong. We cannot evaluate the decision because the criteria are not public. That is the problem. Not the decision. The opacity.
States like Oklahoma are building governance capacity through targeted legislation: chatbot safety, data privacy, well repurposing, study committees. This is necessary work. But state-level governance cannot address the question of who controls frontier AI capabilities at the national scale. That requires federal institutions that do not yet exist, built by a Congress that has not yet acted, informed by a public discourse that has not yet matured.
The thinking machines are already here. The question of who controls them is being answered every day, by default, in the absence of deliberate choice. Every day that passes without governance infrastructure is a day the answer gets harder to change.
Originally published at Structured Emergence, May 5, 2026.