Back in February, Pete Hegseth called Anthropic a supply chain risk. He said it within hours of President Trump telling federal agencies to drop the company's AI. That was supposed to be the old fight, settled and shelved. It's back.
This round kicked off with a directive on export controls. It ended with Anthropic shutting off access to two of its models, Fable 5 and Mythos 5. According to the Wall Street Journal, the order didn't fall out of the sky. It traces back to a piece of security research from Amazon, plus a direct conversation between Amazon CEO Andy Jassy and the White House, as The Verge reported.
Amazon's paper claims researchers coaxed Fable 5, through a chain of prompts, into producing material that could feed a cyberattack. Jassy passed the findings along. Soon after, the government moved to bar foreign nationals from using the model.
Which is where it gets weird.
When the rule locks out the people who built it
A lot of Anthropic's own researchers were born outside the United States. So the ban meant some of them couldn't touch the product they'd spent their careers building. Picture showing up to work and being told the thing on your desk is now off-limits to you, specifically. That's what the directive created.
Anthropic pushed back hard on the framing. The company rejected the idea that what Amazon found was a jailbreak at all, arguing the same weaknesses show up in other models you can already grab off the shelf, GPT 5.5 included. The point being: if this is a security hole, it's an industry-wide one, not an Anthropic problem.
Some outside experts seem to agree. Katie Moussouris, who runs LutaSecurity, wrote on BlueSky that she'd read the paper and "It's not a jailbreak." That's a useful voice to have in your corner. Moussouris is a known quantity in vulnerability research, not a partisan.
The sharper read came from Kate Koren, a former Commerce Department official, who told the Journal the administration's distaste for Anthropic might have colored the call. Hard to prove. Hard to dismiss, too, given the history.
A grudge with a paper trail
This feud was never really about one cybersecurity document. Anthropic has refused to let its models run mass surveillance of Americans or pilot lethal autonomous weapons. Those refusals put it crossways with an administration that wants AI on its own terms.
For a while it looked patched up. After the February blowup, the two sides cooled off and even teamed up to widen access to Mythos. A truce, of sorts. Then the export directive knocked it over, and the rhythm of the past year reasserts itself: cooperation, then a flashpoint, then a fresh round of attacks from Hegseth and other White House allies.
Amazon, for its part, hasn't said a word. That silence is doing some heavy lifting. Jassy's company is a major backer of Anthropic, having poured billions into it, so it's odd to watch Amazon's research become the hinge for a government move that locked Anthropic staff out of their own tools. Either the security concern is real enough to override the investment relationship, or the politics found a handy document and ran with it. Maybe both.
I'd lean toward both, if you're asking. Security papers and political grudges aren't mutually exclusive. A genuine finding can still get weaponized by people who were already hunting for a reason.
The export-control framing matters more than it sounds. Once a model gets tagged as a national security concern and foreign nationals are barred from it, you're no longer arguing about prompts and outputs. You're arguing about who counts as a risk. And a company full of foreign-born researchers becomes collateral in its own product launch.
Watch whether the ban sticks, and whether other labs catch the same treatment. If GPT 5.5 really shares the vulnerabilities Anthropic described, a directive aimed at one company starts to look pretty selective. And selective enforcement of AI export rules, with a defense secretary already calling a major firm a supply chain risk, is the kind of thing that tends to land in court before it gets sorted out anywhere else.