Five days into the Trump administration’s sudden freeze on access to Anthropic’s most advanced AI models—Fable 5 and Mythos 5—and I’m still trying to figure out what, exactly, we’re meant to believe is happening here. Is this a narrowly tailored national security move? Or the opening shot in a much bigger attempt to put the entire U.S. AI sector under an export-control style leash?
Officially, the explanation is “national security.” Unofficially, the way it was done—the speed, the scope, the legal scaffolding that doesn’t quite match the public justification—has turned the decision into a Rorschach test. Depending on who you ask, it’s either overdue caution or an improvised crackdown looking for a clean rationale after the fact.
What the government says it’s doing
Here’s the version of events the administration’s defenders point to. Federal officials, including Treasury Secretary Scott Bessent, were reportedly briefed last week after Amazon CEO Andy Jassy raised alarms about internal research suggesting Anthropic’s Fable 5 could identify weaknesses in Amazon’s cybersecurity software. Amazon, of course, has poured billions into Anthropic over the past three years and supplies some of the GPUs that help power the Claude family of models.
Then came the hammer. On Friday, Commerce Secretary Howard Lutnick sent a letter to Anthropic CEO Dario Amodei stating the company must obtain an official government license before Fable 5 or Mythos 5 can be used by any “foreign persons,” inside or outside the United States. And “foreign persons,” as framed here, doesn’t just mean overseas customers—it reportedly includes Anthropic’s own employees.
If you’re thinking, “That sounds like a paperwork nightmare,” you’re not wrong. Anthropic was told it had less than ninety minutes to comply. The result was predictable: confusion, frantic scrambling, and then the bluntest possible solution—cut off access for everyone to both models.
The letter that didn’t mention cybersecurity
What keeps nagging at me is what wasn’t in Lutnick’s letter. There was no reference to the Amazon briefing, no mention of cybersecurity testing, no “here are the specific risks we observed.” Instead, the letter invoked the Department of Commerce’s Export Administration Regulations (EAR), which allow the government to restrict U.S.-origin products or services if they pose an “unacceptable risk” of use in—or diversion to—military-intelligence end uses or end users. The warning was stark: noncompliance could trigger immediate criminal and civil penalties.
This matters because the EAR framing changes the story. It’s not “your model can find vulnerabilities, therefore it’s dangerous.” It’s “your model may be effectively treated like a controlled export.” That’s a much bigger claim, with much bigger implications, and it helps explain why people across cybersecurity and tech policy are now picking through the details like forensic accountants.
The backlash from cybersecurity experts
After news spread about Jassy’s meeting with federal officials, many cybersecurity professionals rushed to defend Anthropic’s position. An open letter published Sunday to Lutnick and U.S. National Cyber Director Sean Cairncross argued the ban is dangerously misguided.
The central point is one you can summarize in a sentence: finding vulnerabilities is not a unique “Anthropic capability.” According to the letter, many signatories regularly use other foundation models and open-source systems for security testing and red-teaming every day. If the ability to help identify weaknesses becomes the new forbidden threshold, the logic doesn’t just catch Anthropic—it catches almost everybody.
Anthropic echoed that argument in a blog response, adding an even more uncomfortable implication: if “can help discover security flaws” becomes the technical red line that model developers cannot cross, then the practical outcome could be to make it legally perilous for U.S. companies to release new advanced models at all. And in the meantime, the blanket restriction also blocks American cybersecurity defenders from using the models for legitimate defensive work—an irony the open letter emphasizes.
- Capability overlap: vulnerability discovery isn’t exclusive to Fable 5 or Mythos 5.
- Defensive cost: restricting tools harms U.S. security teams as much as it blocks bad actors.
- Precedent risk: the standard could effectively chill future model releases.
Is this even a valid use of export rules?
Legal specialists aren’t uniformly convinced the government’s logic is as airtight as the threat of penalties makes it sound. Alasdair Phillips-Robins—formerly a senior policy adviser at the Department of Commerce and now a fellow focused on tech and international affairs at the Carnegie Endowment for International Peace—argued publicly that Lutnick’s brief letter is so poorly drafted it may not even clearly restrict access through an API or chatbot interface.
His critique cuts to a key technicality: if users access Fable 5 or Mythos 5 through a digital interface, that may not constitute an “export” in the sense the EAR typically governs. If that’s right, then the administration is stretching export law toward something it was never designed to be: a general-purpose tool for policing AI safety or punishing companies Washington considers irresponsible.
Even if the government genuinely believes the models pose a national cybersecurity risk, Phillips-Robins’ broader argument is that export-control regimes aren’t a blank check. If policymakers want to regulate dangerous AI systems, the cleaner route is the boring one: work with Congress and write laws that explicitly do that.
The bigger play: licensing AI by default
This is where the story stops being just about Anthropic. Unless and until Anthropic challenges the decision in court, the administration could use this rationale as a template: impose a licensing requirement across the sector, not for a specific country list, but for any use by “foreign persons.” That’s not a narrow restriction. That’s a new operating environment.
What makes it more striking is the contrast with past posture. The same political ecosystem that often praised a “hands-off” approach to AI now appears to be flirting with a federal control regime based on export laws. Reporting from The Information this week suggests officials have been exploring precisely that—a system to police model capability and reach using export authorities.
So the question I keep coming back to is one you might be asking too: is this a one-off panic response, or the first test case?
The timing, the G7, and the “digital blockade” effect
The timing adds another layer. Three days after Lutnick’s letter, President Trump arrived in Évian-les-Bains for the annual G7 summit. And while plenty of people hear “national security” and think “Russia, Iran, China,” this restriction isn’t limited to geopolitical adversaries. It’s global.
That means even America’s closest allies are now cut off from two of the most advanced AI models in circulation. In practice, that starts to look less like a targeted safeguard and more like leverage—a kind of digital blockade, where access to frontier AI becomes something Washington can grant, deny, or trade.
Trump has already refused British Prime Minister Keir Starmer’s request for an exemption. If allies can’t get carve-outs, the message is blunt: the U.S. is willing to treat advanced AI like a strategic resource—and to make friends feel the pinch along with everyone else.
Whether you find that prudent or alarming probably depends on what you think the real objective is. But five days in, the silence about specifics—the absence of clear criteria, timelines, or a defined endgame—doesn’t inspire confidence. It invites suspicion. And in a moment when AI policy needs legitimacy as much as it needs القوة, this is a strange way to build either.
Comments