Last week, Anthropic was forced to pull its Fable 5 and Mythos 5 models offline for all users—including US-based employees and foreign nationals inside the country—after an abrupt executive order citing “national security authorities.” The directive, still lacking a public legal rationale, marks the first time US export controls have been applied to restrict access to an AI model delivered purely as a hosted service.

Export control frameworks were built for physical goods: weapons, hardware, then software and source code. Model weights—the trained parameter files that can be copied and run elsewhere—were added under the Biden administration, then dropped by Trump. But Anthropic’s case doesn’t fit that mold. No weights or source code changed hands. Users query a remote API and get back text. The “export” would have to be either the model’s output or the act of access itself—a known gap that Congress is now trying to close via legislation moving through the Senate.

Georgetown’s Hanna Dohmen called it “an open question” whether the order stretches existing rules. UC Berkeley’s Andrew Reddie noted that while arms and export regulations give the government wide latitude, “equivocation by successive administrations regarding what the responsibilities of model developers are” leaves firms guessing. If the target was capability, every frontier lab should be worried. If it was a specific jailbreak concern, the government needs to define what safeguards suffice. If it was political, the system is broken.

The core engineering takeaway: this is not a sustainable governance model. Ad hoc, opaque interventions create compliance uncertainty for any team deploying large language models via API. If the US wants to maintain leadership, it needs clear, predictable rules—not a shutdown order delivered without explanation.

Source: The Verge