The Trump Administration Is Lifting Its Export Controls on Anthropic's Mythos and Fable AI Models
The Commerce Department's decision to lift export controls on Anthropic's Mythos 5 and Fable 5 models isn't a victory for innovation. It's a case study in how AI policy gets made when the adults leave the room and the dealers take over the table.
Secretary Howard Lutnick's letter to Anthropic cofounder Tom Brown — "A license is no longer required for the export, reexport, or in-country transfer... of the Mythos or Fable models" — reads like a hostage release note. And in a sense, it is. The hostage was American AI competitiveness. The ransom was a promise to "proactively detect and address security risks" that Anthropic itself had previously declared impossible to fully eliminate.
The Pivot That Tells You Everything
Rewind six months. Anthropic's position was principled, even brave: you cannot guarantee zero jailbreaks. The company argued that the administration's demand for airtight security against capability extraction was a technical fantasy. Dario Amodei, Anthropic's CEO, said as much in meetings with officials. He treated the conversation as a technical negotiation between equals.
That approach failed. Amodei got replaced in those meetings by Tom Brown — not because Brown is more knowledgeable, but because, as WIRED reported, "officials liked more on a personal level."
Let that sink in. The fate of the world's most powerful AI models hinged on whether a Cabinet secretary vibed with the messenger.
Anthropic then did what every company does when principle meets power: they folded. They stopped relitigating the conceptual issue and started building "more robust safeguards" — industry speak for we'll try harder to stop the thing we told you couldn't be stopped. The administration got what it wanted: a willing partner who would pretend the unsolvable is solvable.
Export Controls as Theater
The Mythos and Fable restrictions were always performative. The Biden administration pioneered AI export controls with the October 2023 executive order and subsequent Commerce rules targeting advanced compute and model weights. The logic was coherent: keep frontier capabilities out of adversary hands.
But the Trump administration inherited a framework it didn't build and doesn't understand. Lutnick and national cyber director Sean Cairncross aren't technocrats; they're dealmakers. They see export controls not as national security architecture but as leverage — chips to trade for concessions, headlines, and the appearance of action.
Lifting controls on Mythos 5 — the model Anthropic itself considered too dangerous for wide release — suggests the threat model was never the point. The point was the deal.
The Safety Theater Problem
Anthropic's new commitments are worth examining. "Proactively detect and address security risks." "Work diligently on protocols and standards." These are process promises, not outcome guarantees. They create the appearance of governance without the substance.
We've seen this movie before. Voluntary commitments from OpenAI, Google, Microsoft — the 2023 White House pledges, the Frontier Model Forum, the endless red-teaming exercises. None have prevented jailbreaks. None have stopped capability proliferation. The only thing that has slowed dissemination is compute scarcity, and that moat is shrinking.
The administration knows this. Anthropic knows this. But both need the theater. Anthropic needs to sell models globally. The administration needs to claim it "secured" AI. So they sign a paper that changes nothing technically but changes everything politically.
What This Means for the Rest of Us
The precedent is corrosive. Every frontier lab now knows the playbook: resist initially, then concede on process, swap executives if needed, promise safeguards you can't deliver. The government will blink. It always blinks when the alternative is admitting it has no actual regulatory framework.
Congress has abdicated. The EU AI Act moves at regulatory geology speed. The states are fragmented. So we get governance by press release and personal chemistry.
Mythos 5 and Fable 5 will now flow legally to allies, partners, and — inevitably — adversaries through cutouts, cloud APIs, and weight theft. The safeguards Anthropic builds will be bypassed within weeks; the cat-and-mouse game favors the mice. But the paperwork will be in order.
The Real Cost
The cost isn't just that powerful models escape. It's that we've replaced policy with relationship management. The next Anthropic — and there will be a next one — won't even bother with the initial resistance. They'll lead with the concession. They'll staff their government affairs teams with people Lutnick likes.
That's not AI safety. That's regulatory capture at hyperspeed.
Lutnick's letter closes a chapter. But the book — a coherent American strategy for frontier AI — remains unwritten. And nobody in this administration seems interested in writing it.