Episode Details

On June 12, 2026, the Bureau of Industry and Security told Anthropic that it needed a license before any foreign person, anywhere in the world, could access its Mythos and Fable artificial intelligence models. The directive did not name the weights beneath those models or the chips that trained them as the controlled item. It named the model itself. Every prior artificial intelligence export rule, going back to the January 2025 Framework for Artificial Intelligence Diffusion, had controlled model weights as the regulated item under Export Control Classification Number 4E091. The June 2026 directive was the first time Commerce reached past the weights to the model as a controlled item in its own right, and the popular account of artificial intelligence export policy, built almost entirely around chip sales to China, missed it entirely.

Most coverage of artificial intelligence export controls still describes a fight over graphics processing units. Nvidia’s H100 and H200 chips headline the story because a Houston based smuggling operation named Operation Gatekeeper moved 160,815,000 dollars of them toward buyers connected to the People’s Republic of China before the Department of Justice announced its case on December 9, 2025. That story is accurate and it is also incomplete, because it treats the export control regime as a hardware problem when the June 2026 Anthropic directive shows Commerce now treats the software running on that hardware as the thing worth controlling.

The gap widened further because the rule that was supposed to govern this territory does not currently exist in enforceable form. The Biden administration’s Framework for Artificial Intelligence Diffusion took effect January 13, 2025 with a compliance date of May 15, 2025, sorting the world into three tiers and assigning per country compute allocations. The incoming administration’s Bureau of Industry and Security announced on May 13, 2025 that it would stop enforcing that rule, calling its tier system a diplomatic liability, and said a replacement rule would follow. As of this writing, no replacement has been published. Commerce reached Anthropic through the underlying statute directly, not through a successor regulation, which means the rule a second company would need to comply with does not yet exist in public text.

The mechanism matters more than the headline. Section 1758 of the Export Control Reform Act of 2018, codified at 50 U.S.C. 4817, gives the Secretary of Commerce authority to identify and control emerging and foundational technologies. The Bureau used that authority in the January 2025 rule to define a controlled artificial intelligence item as the model weights of a closed weight system trained past ten to the twenty sixth power computational operations. A closed weight model is one whose weights are never published; an open source research model released for publication sits outside that entry by design. The June 2026 directive to Anthropic cited Section 744.22(b) of the Export Administration Regulations and the same statutory provision, but it named the model as the controlled item rather than the weights. That distinction carries real consequences. A license tied to model access reaches every deployment surface where a foreign national could interact with the system, and a physical file transfer is no longer the only trigger.

The Government Accountability Office settled a related question on May 12, 2026 that deserves more attention than it received. In decision B-337935, the Comptroller General found that the Bureau’s May 2025 press release announcing non enforcement of the diffusion framework was itself a rule under the Administrative Procedure Act, one Commerce never submitted to Congress as the Congressional Review Act requires. The opinion rejected the idea that an agency can quietly retire a licensing requirement through a communications strategy rather than a rulemaking. Read against the June 2026 Anthropic d