Export controls are laws that govern which countries and individuals can receive sensitive technology. For decades, the US applied them to nuclear materials, weapons, and semiconductors. Governments are now extending these rules to frontier AI models — and in June 2026, Anthropic’s users worldwide learned what that looks like in practice.

What Are Export Controls?

Under the Export Administration Regulations (EAR), the US Commerce Department’s Bureau of Industry and Security (BIS) administers controls on “dual-use” goods — civilian technology that could also serve military or national-security purposes. Each controlled item receives an Export Control Classification Number (ECCN); certain destinations require a license, or face an outright ban.

Export controls are not new. They have long applied to encryption software, aerospace components, and nuclear materials. The underlying principle is that some technology is too sensitive to share without restriction — an adversary could use it to build weapons, close a strategic gap, or destabilize the global order.

How AI Fell Under Export Law

The first major extension of export controls to AI came through hardware. Starting in 2022, the Commerce Department restricted the sale of NVIDIA’s A100 and H100 GPUs — the chips that power frontier AI training — to China and other adversary nations. The logic: whoever controls the hardware that trains AI shapes who develops next-generation military AI.

But chips are physical objects. AI software is different. A frontier model’s “weights” — the billions of numerical parameters that encode what the model has learned — can be transmitted as a file. This created a new regulatory question: if you control the hardware that trains AI, should you also control the model it produces?

The AI Diffusion Rule — and Its Reversal

In January 2025, the Biden administration answered yes. The “Framework for Artificial Intelligence Diffusion” (Federal Register doc 2025-00636) introduced the first US export controls on AI model weights, along with a global three-tier country system:

  • Tier 1 (18 close allies — Australia, Canada, France, Germany, Japan, South Korea, the United Kingdom, and others): essentially unrestricted access to advanced AI chips and models.
  • Tier 2 (most of the world, including much of Europe not in Tier 1, India, and Southeast Asia): access subject to per-country compute caps and licensing requirements.
  • Tier 3 (China, Russia, Iran, North Korea, and other arms-embargoed nations): near-total restrictions.

The rule also applied to the weights of frontier AI models themselves — creating classification ECCN 4E091 for models trained with more than 10²⁶ computational operations. Open-weight, publicly released models were explicitly exempted.

The Trump administration rescinded the rule on May 13, 2025, two days before it was set to take effect, citing competitive concerns and diplomatic friction with allied nations downgraded to Tier 2. No comprehensive replacement had been published as of mid-2026, though the underlying chip controls on China remained in place.

When a Jailbreak Triggered a Global Shutdown

The Fable 5 incident in June 2026 showed what AI export enforcement looks like when applied to a deployed cloud model.

Three days after Anthropic launched Claude Fable 5 and Mythos 5, the Commerce Department issued Anthropic a directive: suspend access for any foreign national, worldwide — including foreign nationals employed at Anthropic. The trigger: Amazon researchers had reported a jailbreak technique that, the government argued, could turn the models into unrestricted cyber tools.

Because Anthropic had no reliable way to verify every user’s nationality in real time, it took the only compliant path available: it disabled Fable 5 and Mythos 5 globally. The shutdown lasted approximately three weeks. After Anthropic deployed a new safety classifier that blocks the jailbreak technique in more than 99% of cases, and after the Commerce Department’s own researchers validated the result, the US lifted the export controls on July 1, 2026. Fable 5 was restored globally; Mythos 5 remained restricted to a vetted group of US critical-infrastructure organizations.

The episode demonstrated two things. First, a cloud-delivered AI model can be treated as a controlled export — no physical handoff required. Second, the US government can effectively take a frontier AI model offline worldwide, with little warning.

What This Means for Users and Businesses

For users in Tier 1 allied countries, AI export controls are mostly invisible under normal conditions. The practical risk is sudden disruption — a service going dark without notice, as the Fable 5 case showed.

For businesses building products on US AI APIs, the incident highlights that geopolitical risk is also operational risk. A government directive can interrupt service regardless of where customers are located. Contingency planning — alternative providers, graceful degradation, clear communication protocols — is increasingly standard practice.

For AI developers, the case sets a precedent: if a model is judged to present a national security risk, regulators can compel a global shutdown, not just restrictions in specific countries.

In the News

The US lifted export controls on Fable 5, and Anthropic began restoring access globally. Read the full story: Anthropic Restores Fable 5 Globally After US Lifts Export Controls.

FAQ

What is the Bureau of Industry and Security?
BIS is the US Commerce Department agency that administers export controls on dual-use civilian technology. It can restrict which countries, companies, and individuals are permitted to receive controlled US technology — including advanced AI chips and, under some rules, AI model weights.

Do export controls apply to open-source AI models?
The Biden-era AI Diffusion Rule (now rescinded) exempted openly published model weights entirely. The June 2026 enforcement action targeted a specific proprietary model, not open-source software. Under current rules, openly published weights are generally harder to control, though the policy landscape remains unsettled.

Can a government shut down an AI service in my country?
Yes, indirectly. If the AI company is US-based, the US government can order it to suspend access for foreign nationals or specific destinations. The company must enforce that restriction — or, if it cannot verify user nationality in real time, take the service offline entirely. Anthropic’s response to the Fable 5 directive is the clearest example on record.

What happened to the Biden AI Diffusion Rule?
The Biden administration published the Framework for Artificial Intelligence Diffusion in January 2025. The Trump administration rescinded it in May 2025, citing competitive concerns and allied-nation objections to Tier 2 classification. The underlying chip controls on China remain in force.