Anthropic did everything the federal government asked. It let reviewers examine its Mythos model, incorporated their feedback, and installed guardrails against high-risk uses in cybersecurity and bioweapons. Within days of releasing Claude Fable 5, the Commerce Department imposed export restrictions—reportedly after Amazon CEO Andy Jassy phoned the White House.
The Trump administration’s executive order on AI, issued just before Fable’s launch, encouraged companies to voluntarily give the government up to 30 days of access to their “covered frontier models.” But that term remains undefined. The order tasks the Treasury secretary, Defense secretary, Homeland Security secretary, the science adviser, and the Commerce secretary with setting the threshold within 60 days—a process that hasn’t even begun.
Even if the bureaucracy aligns, 30 days is far too short to evaluate probabilistic models that evolve with use. A model that appears safe during a month of testing could later enable harm. Government reviewers, meanwhile, face staffing and expertise gaps. As one former Homeland Security official noted, when frontier model companies sought similar sign-offs in 2024, his office refused—citing the same constraints of time, bandwidth, and post-review model drift.
The executive order revives that flawed logic—and adds a new danger: if a government-cleared model later causes harm, the blame will shift to Washington, not the company. That’s a political liability no administration wants.
Anthropic’s experience shows why no company will trust this process. It cooperated before the order required it. The government reviewed the model. Anthropic built in safeguards. Then Commerce restricted it—not because of any review finding, but because a rival CEO made a call. Why would any firm submit to a voluntary process that offers no legal protection and invites arbitrary reversal? In fact, submitting may draw more scrutiny than staying quiet. As Fable proves, the government can always reach for export controls regardless of what a review says.
A better approach exists. The Frontier Model Forum already facilitates private-sector information sharing on risks. Building on that—with continuous engagement between government and AI developers—would track model evolution, build government expertise, and create relationships that encourage cooperation rather than fear. The lesson of Fable isn’t that cooperation failed; it’s that the current framework is a trap.
Meanwhile, as debates over AI regulation intensify, other pressing issues—like the public’s strong support for birthright citizenship or the Supreme Court’s unanimous ruling on gun rights for marijuana users—show how quickly policy can shift. And as two-thirds of Americans now see voting rights under threat, the stakes for transparent governance have never been higher.
