When The State Controls The Switch

Welcome to Memorandum Deep Dives. In this series, we go beyond the headlines to examine the decisions shaping our digital future. 🗞️

This week, two of the most advanced AI systems on the market vanished in a single day. No chips were seized, no research was banned, and no court issued an order. One government directive was enough, and the consequences reached users on every continent.

The trigger was a dispute over a jailbreak, and that is the part most coverage has fixated on. But the jailbreak is not the story. The story is the instrument that the episode quietly exposed, one that has been forming for years while everyone watched the chips.

For most of its history, AI sat outside the machinery governments use to control strategic technology. That is no longer true, and the way this shutdown happened says far more about the next decade than the argument that set it off. The mechanism deserves a closer look.

How frontier AI became a strategic resource

On June 12, a single government directive took down the two most capable AI models on the market, affecting users worldwide. No chips were seized, and no research was banned. The state simply reached into a live commercial service and decided who was allowed to use it. That mechanism, not the jailbreak that triggered it, is the development worth tracking.

For decades, the United States approached strategic technologies through the logic of control. Nuclear materials, advanced weapons systems, semiconductors, telecommunications infrastructure, and critical energy resources were never treated as ordinary commercial products because access to them could shape economic power, military capability, and geopolitical influence far beyond the organizations that produced them. The purpose of export controls was therefore not simply to regulate trade. It was to determine who could access capabilities that governments considered strategically significant.

Artificial intelligence has spent most of its existence outside that framework. Software has traditionally been easier to distribute than hardware, easier to replicate than physical infrastructure, and far more difficult to contain within national borders. Even as AI models became more powerful, the prevailing assumption was that governments would focus their attention on the underlying ingredients required to build them, particularly advanced chips and the computing infrastructure needed to train frontier systems.

However, that assumption has steadily eroded over the past several years as policymakers have moved higher up the AI stack. Restrictions that initially targeted advanced semiconductors expanded to cover the equipment used to manufacture them, the cloud infrastructure required to train large-scale systems, and eventually the models themselves. The shift became explicit in January 2025, when the Bureau of Industry and Security introduced export controls on certain advanced AI model weights for the first time, signaling that Washington now viewed frontier models as strategic assets rather than ordinary software subject to direct regulation. What remained unclear was how far that logic would ultimately extend.

From controlling hardware to controlling models

That question moved from theoretical to practical when Anthropic announced it was disabling Claude Fable 5 and Mythos 5 after receiving a U.S. government export-control directive prohibiting access by foreign nationals, including those working at Anthropic. Because a model delivered via a cloud interface cannot reliably determine a user's citizenship in real time, the company argued that compliance effectively required shutting down the systems for everyone. Other Claude models remained online, but the two most capable systems disappeared overnight.

At the time the controls were announced, the Commerce Department believed a method existed to bypass certain safeguards in Fable 5, potentially unlocking advanced cybersecurity capabilities.

However, Anthropic contested that assessment, arguing that the demonstrated vulnerabilities were narrow, already known, and comparable to capabilities available through other publicly accessible models. The company further shared that recalling a model used by hundreds of millions of people over a narrow jailbreak would, if applied across the industry, "essentially halt all new model deployments."

The clash over export controls is the latest chapter in a wider confrontation between Anthropic and the U.S. government. Earlier disputes over the company's refusal to permit Claude's use in autonomous weapons and domestic surveillance led federal officials to characterize Anthropic as a 'supply chain risk', setting off a series of legal battles that remain unresolved.

The unresolved question underneath every one of these fights is the same: where a legitimate security interest ends and a competitive one begins, when the same lever serves both.

The first test of AI access control

Regardless of the disagreement, what makes the episode significant is how it occurred, more than the intervention itself.

The government did not restrict future research, block access to training chips, or impose conditions on a forthcoming release. Instead, it reached directly into a commercial AI service that had already been deployed and effectively determined who could and could not access its capabilities.

When viewed through that lens, the Anthropic shutdown looks like evidence that the object of control is changing, rather than a one-off dispute over a jailbreak. For years, export policy has focused on the hardware required to create advanced AI. The June directive suggests that policymakers are becoming increasingly interested in governing access to the capabilities themselves.

The strongest objection is that nothing structurally new happened here. Governments have always held leverage over services that touch their soil, from telecom wiretap mandates to financial sanctions enforced through the dollar clearing system. A one-time directive answered under protest is not the same as a durable regime, and Anthropic is in active litigation precisely because the boundaries remain contested. By this reading, the shutdown is a jurisdictional skirmish, not a phase change. What weakens that objection is repeatability: the same instrument can be aimed at any hosted model, by any state with jurisdiction over the provider, without new authority or public debate.

The rise of AI as a service

The underlying reason for this shift becomes clearer when viewed through the lens of how modern AI systems are delivered. Traditional software could be purchased, downloaded, installed, and operated independently of its creator. Once distributed, control largely passed from the developer to the user. Frontier AI works differently. The most capable models increasingly exist as services rather than products, hosted in data centers and accessed through application programming interfaces. Users do not possess these systems in any meaningful sense. They interact with them remotely, drawing on capabilities that remain under the provider's control.

That seemingly technical distinction has important geopolitical consequences. Because access is mediated through infrastructure controlled by a relatively small number of companies, governments no longer need to prevent the transfer of a technology to regulate its use. They can instead influence who is allowed to access it, under what conditions, and for what purposes. As a result, the strategic question shifts from who owns the technology to who is allowed to use it.

Who decides who gets intelligence?

In many ways, this follows the same logic that has historically driven export controls. Semiconductors did not become strategically important because of the silicon itself. They became strategically important because modern economies, military systems, and critical infrastructure increasingly depended on the computing power they enabled. Frontier AI is undergoing a similar transformation. As advanced models become more capable in areas such as software engineering, cybersecurity, scientific discovery, and military planning, policymakers are beginning to view access to those capabilities through the same lens they once applied to advanced chips: as a strategic resource rather than an ordinary commercial service whose distribution can shape economic and national power.

The cleaner way to hold this is to stop thinking of frontier models as products and start thinking of them as metered utilities. You do not own the electricity in your wall; you are granted access to it under conditions, and that access can be cut at the meter. Frontier AI is converging on the same arrangement, except the meter now sits inside a policy decision rather than a billing one.

The mechanism revealed in June will not disappear when this dispute ends. Any government with jurisdiction over an AI provider can use it without passing new laws, largely outside public view. The question is no longer whether such powers will be used again, but who gets to decide the conditions of access when the state controls the switch and the technology serves everyone.