Anthropic released its two most advanced AI models on June 9. Three days later, the Trump administration used export controls to force them to pull them back.
Weeks later the company's most capable models still aren't available. The standoff is now a fight over a guarantee that experts say no AI lab can give.
Anthropic has spent roughly two weeks in a standoff with the White House over its most advanced AI models, and nothing is resolved. The company released Mythos 5 for trusted organizations on June 9, along with a public version called Fable 5. Three days later, the government imposed export controls, the same kind of trade rules that restrict sending sensitive technology abroad, and forced the models offline for all users.
The trigger was a security scare. Amazon CEO Andy Jassy, whose company is one of Anthropic's biggest investors, called Treasury Secretary Scott Bessent directly to flag an alleged way to jailbreak the models' guardrails, meaning a trick that gets an AI to ignore its safety rules. The NSA reviewed it and the Commerce Department acted. A letter from Commerce Secretary Howard Lutnick to Anthropic CEO Dario Amodei, published by Bloomberg on June 15, required a license to export the models "to all destinations worldwide and to all foreign persons."
The administration's position has since hardened. Officials told Wired that getting Fable 5 back is Anthropic's problem to fix, and that the company must guarantee its guardrails cannot be jailbroken at all, something most experts say is simply not possible.
On Episode 221 of The Artificial Intelligence Show, SmarterX founder and CEO Paul Roetzer examined what the fight reveals about U.S. AI policy.
2018 - Export Control Reform Act used to force the models offline
June 9 - release date for Mythos 5 and Fable 5
3 - # of days after their release that the two Anthrophic models were forced from the market
7 months - how far ahead of past trends Epoch AI found Mythos in exploit development
100% - the unbreakable-guardrail guarantee the government is demanding
Roetzer's notes the administration is doing the opposite of what it promised 15 months ago.
The impossible ask. The government wants a guarantee no lab can make: an impentetrable AI model. As Roetzer put it, guardrails can be circumvented in every model, not only for Fable 5. "It's not how these things work. It's not like patching a bug in software that you fix." Either the government does not understand how AI models work, he says, or officials know the demand is impossible and are putting the information out there for other reasons.
The standard contradiction. In February 2025, weeks into the new term, Vice President JD Vance gave a widely praised speech at the AI Action Summit in Paris. He warned against incumbents using safety rules to box out rivals and declared, in his prepared remarks, that "the AI future is not going to be won by hand-wringing about safety. It will be won by building." Vance also said the administration would ensure "American AI technology continues to be the gold standard worldwide."
Roetzer flagged the gap directly:
"Keep in mind, they want to be the standard, and yet they're not allowing foreign actors to use the model. They're the ones that stepped in and did this. When you're limiting access to your models, it becomes very hard to be that standard."
—Paul Roetzer, founder and CEO, SmarterX, Episode 221 of The Artificial Intelligence Show
The cyber reality. The capability concern is real. Epoch AI reviewed the public evidence and found Mythos Preview was a large improvement in exploit development, much better than GPT-5.5 and about seven months ahead of past trends. And Mythos 5 was modestly better still. A widely shared claim that Mythos "broke into almost all" NSA classified systems "in hours" came from The Economist, but the journalist who wrote it later walked it back, saying it was missing caveats and should not be read literally. Roetzer noted he had found that line odd when the article first ran.
So something changed. Either the administration did not understand scaling laws, the well-known pattern that bigger models keep getting more capable, or its position shifted. "It's really hard to look at what the government is doing right now and think they actually know what they're doing," Roetzer says, "otherwise they wouldn't have said what they said 15 months ago."
The uncomfortable lesson for business leaders is vendor concentration risk. If a single government directive can pull a frontier model offline overnight, then building a critical workflow on one AI provider is a single point of failure that has nothing to do with the technology and everything to do with politics. This is no longer a hypothetical. Anthropic's best models have been dark for weeks while the IPO clock ticks.
The government regulations might also not be applied evenly. Every future model from Google and OpenAI almost certainly has similar cyber capabilities, yet Roetzer sees signs of preferential treatment. At the G7 summit, Demis Hassabis and Sam Altman flanked President Donald Trump while Dario Amodei sat at "the kids' table." OpenAI president Greg Brockman is a large donor to a pro-Trump super PAC or Political Action Committee. Roetzer thinks OpenAI may get preferential release approval over Anthropic, which he called a "very slippery slope."
The thaw may already be starting. Trump told Axios he no longer sees Anthropic as a national security threat and that the company "behaved very responsibly,." He told The Wall Street Journal that negotiations with the company are "going fine." Even so, the formal Commerce directive still stands, the models remain offline, and Anthropic says only that they will be back "in coming days," all while the company moves toward what could be the largest IPO ever.
A bigger ban could be next. Roetzer predicts the administration could move within about 30 days to ban U.S. companies from using Chinese open-source models such as DeepSeek, which firms are increasingly adopting to cut costs. If that happens, the export-control fight stops being about one company and becomes a broader struggle over which AI models American businesses are allowed to use.
When the rules around AI can change in 72 hours, the organizations that stay ahead are the ones with governance already in place. Yet according to the 2026 State of AI for Business Report, only 13% of organizations have all four governance foundations in place to scale AI responsibly: an AI roadmap, an AI council, generative AI policies, and an AI ethics policy. A full 32% have none of them.
That gap matters most in moments like this one, when a model your team depends on can be pulled by a government directive overnight. Built on more than 2,100 responses, the report maps where companies actually stand on adoption, policy, and readiness, and what leaders are doing differently. Read the full report →