If export controls are an AI kill switch, the process needs reform

Skip to content

The views expressed by contributors are their own and not the view of The Hill

The administration’s June 12 export control order to Anthropic effectively forced the company to recall its most powerful AI models — Mythos 5 and Fable 5. Government officials seemed to think these two AI models could enable attacks on U.S. critical infrastructure that were otherwise unavailable to adversaries. So, they issued an order that instructed Anthropic not to make them available to foreign companies or to foreign nationals, even if they were in the U.S.

Anthropic could not distinguish access to its models by nationality or citizenship status and had to withdraw the models from public availability in order to comply with the order. A few days later, OpenAI also limited the release of its latest model, GPT-5.6, at the administration’s request, apparently fearing an export ban if it did not comply.

After a few weeks, the administration partially lifted its Mythos 5 ban, allowing the model to be released to a selected group of U.S. users. On June 30, the administration wrote a letter to Anthropic saying a license would no longer be required for Mythos or Fable. Anthropic released Fable to the general public but wrote that it continued to limit Mythos access to “select U.S. organizations.” On July 8, the administration approved a “wide release” of OpenAI’s GPT-5.6.

Congress intended export control orders to fulfill a variety of purposes connected to foreign policy and national security. But in this case, government officials clearly intended the export order to compel a consumer product recall.

The administration apparently first asked Anthropic to voluntarily remove the model from circulation and then coordinate with the government to address security vulnerabilities. Export controls, said one official, were a “last resort.” Anthropic said they were given “90 minutes to take the model down.”

The result, as one person familiar with the incident said, is “a de-facto licensing regime.” And there is no guarantee of continued approval. As Martin Chorzempa, a senior fellow at the Peterson Institute put it, “You have no idea whether the U.S. government is just going to shut off your access to any future models.” Even after restoring access to Fable and Mythos, the administration made it clear that it retained the right to reimpose a license requirement.

Jessica Tillipman, associate dean for government procurement law at George Washington University, writes that the dispute over whether Anthropic’s new models posed a significant cybersecurity threat is “precisely what a fair process is meant to resolve, not a 90-minute ultimatum.”

Of course she is right. But law professor Alan Z. Rozenshtein may also be right in thinking that the underlying Export Control Reform Act of 2018 allows the government to adopt this “kill switch for frontier AI models” without due process. While there are still questions about whether providing access to an AI model is legally an “export,” the export control law clearly authorizes Commerce Department officials to issue secret, binding “is informed” letters to companies saying they need an export license for certain items.

By an explicit provision of the statute, the export control regime is largely exempt from the public notice and comment rulemaking requirements of the Administrative Procedure Act, which also means that injured parties cannot take the government to court claiming that the government’s actions were arbitrary and capricious.

To misuse its export control authority to suddenly create an AI model kill switch that no one suspected even existed is a paradigm case of the arbitrary and capricious regulatory action that the Administrative Procedure Act was invented to prevent. The very existence of this kind of untrammeled authority undermines U.S. tech companies abroad, who cannot now credibly pledge to maintain AI model access to their customers.

A centrist French presidential candidate recently described this secret authority over AI models as America’s Strait of Hormuz. Until some transparent, accountable process is wrapped around these export control decisions, U.S. AI companies are at risk of being perceived as unreliable business partners around the world.

As Congress considers regulating AI to reduce consumer and catastrophic risks, it might also consider reforming the export control statute so that the administration cannot simply impose an export control without public notice and comment as it did in the Mythos-Fable case and in other cases involving chip export controls. Affirmatively applying Administrative Procedure Act protections might be one easy way to go.

In addition, since the way the administration used the statute makes it equivalent to a product recall, the rules governing recalls for the National Highway and Traffic Safety Administration and the Consumer Product Safety Commission might provide a model for policymakers to consider.

Some substantive standards for imposing export controls on AI models might be necessary as well, but the key thing is to require the government to prove its case publicly and provide accountable review to prevent arbitrary and capricious action.

In a way, the administration has done the public a favor. By exposing the dangerously open-ended authority the export control statue provides, it has made the case for reform crystal clear. Congress must act now.

Mark MacCarthy is the author of “Regulating Digital Industries,” an adjunct professor at Georgetown University’s Communication, Culture and Technology Program, a nonresident senior fellow at the Institute for Technology Law and Policy at Georgetown Law and a nonresident senior fellow at the Brookings Institution.

Tags

Administrative Procedure Act

ai kill switch

AI models

AI regulation

Anthropic

Anthropic ban

Anthropic's Mythos

congress

export controls

Fable 5

Fable 5 model

GPT 5.6 release

GPT-5.6

Mythos 5

Mythos 5 model

national highway and traffic safety administration

OpenAI

Strait of Hormuz

U.S.

U.S. Consumer Product Safety Commission

Copyright 2026 Nexstar Media Inc. All rights reserved. This material may not be published, broadcast, rewritten, or redistributed.