Artificial intelligence is becoming a cornerstone of American economic competitiveness and national security. Yet, even though policymakers have devoted significant attention to how AI is used domestically, much less attention has been given to a similarly important question: who should have access to America’s most advanced AI models?

Current AI models are capable of more than just generating text, and they are increasingly assisting with scientific research, cybersecurity, and other advanced problem-solving tasks. For example, Anthropic’s Mythos 5 was designed specifically for advanced cybersecurity and biology research and was initially released only to a small group of vetted organizations due to its capabilities. Because of these skills, some models present substantially greater risks than earlier generations.  Recognizing these concerns, policymakers have begun considering restrictions on exporting the most advanced models to certain countries, such as China and Russia, and to certain organizations. In June 2026, the Commerce Department ordered Anthropic to suspend access to its frontier AI models, Fable 5 and Mythos 5, for foreign nationals under export control authorities. Anthropic, however, argued that the government had not provided sufficient evidence for the restriction. Although the restrictions were lifted several weeks later, this incident highlighted the competing interests at play in this issue, with no transparent procedures. The problem lies in the current framing of the issue as a binary choice between restricting and allowing exports. In reality, AI capabilities and export options span a much broader spectrum, and the policy should reflect that.  

AI model export restrictions are a difficult balancing act. If done too aggressively, and U.S. developers face significant barriers that foreign competitors do not, investment and innovation may shift abroad. However, allowing unrestricted exports could provide American adversaries with technologies that could undermine U.S. interests. The answer is not choosing one extreme or the other. Rather, it is building a strong federal framework that allows responsible access while managing other security risks. Poorly defined restrictions also create uncertainty. Congress has already recognized the strategic importance of AI infrastructure through the CHIPS and Science Act, which invested more than $52 billion in domestic semiconductor manufacturing to reduce dependence on foreign supply chains. AI model policy should similarly protect national security without discouraging American innovation.

Companies need predictable rules to determine which models require review and what compliance obligations exist. Such restrictions also ignore the reality that not every advanced model has the same level of risk. A model trained for creative writing does not require the same level of regulatory oversight as one conducting advanced scientific research. AI export policy needs to be durable and clearly define which models warrant additional oversight. Congress should provide American developers with regulatory certainty and ensure accountability through standardized safety evaluations and reporting. For example, Rep. Moran introduced a bill requiring incident reporting for dangerous models. This policy would combine safety with innovation. 

Ultimately, the conversation surrounding frontier AI models should focus on designing a system that acknowledges varying levels of risk and preserves American leadership in AI innovation. As AI capabilities continue to advance, export policy will become increasingly consequential.