Anthropic’s IPO pitch has a new problem: the government can shut it down

Anthropic has filed confidentially to go public as soon as this fall, pitching investors on a single thesis: that it leads the enterprise AI market and intends to transform the world with it. 

But it has a problem that just won’t go away: the U.S. government becoming its outright adversary. For a company carrying a nearly $1 trillion valuation into a public offering, it has been blacklisted twice now by the federal government, forcing investors to consider whether that mega-valuation is fully pricing in the fact that the government is willing to switch off its flagship product overnight. 

After a shock announcement on Friday that Anthropic would be taking its two newest models, Fable 5 and Mythos 5, offline due to the Commerce Department barring foreign nationals from using it, the feud has not let up. It followed a blacklisting by the federal government over national security concerns in March.

Secretary of War Pete Hegseth took a victory lap, indicating the government will maintain its enmity with Anthropic. “Three months ago, [the Department of War] kicked Anthropic out of our building — forever,” he posted on Sunday. “Every passing day proves why that was the right move.” 

Even as Anthropic sent senior technical staff to Washington over the weekend to argue against the export control, the Department of War wrote on X to say it has moved at least two-thirds of its AI workflows off Anthropic’s models since the two clashed over military use of Claude.The department, it announced Monday afternoon, “will no longer be single-threaded to one AI provider.” Its warfighters would instead get “a diverse suite of AI capabilities to ensure they achieve true decision superiority.”

“Heck yeah there’s regulatory risk,” said David Linthicum, a longtime cloud analyst, who argued that anyone betting on these companies should have seen government intervention coming. Pour trillions into building something powerful enough to scare people, and the government’s interest is clearly part of the deal. He doesn’t expect the standoff to last: within 48 hours, he figures, Anthropic and the administration will kiss and make up—”and then in six weeks it’ll happen again, and Anthropic will have to make another trip down.” 

That reactive style of regulation isn’t just inefficient, he said; it carries a “chilling effect” on research. OpenAI might now think twice before shipping its next model, wary of losing billions in revenue to the same kind of order. And foreign customers may drift toward home-grown options — China has its own; Europe has no real answer.

But Anthropic hurts itself the most by getting into such a battle with the federal government, Lithicum said, not to mention its theoretical future valuation and stock price.
Amazon–an Anthropic frenemy?

It seems strange that the trigger was one of Anthropic’s biggest investors; Amazon, in for some $8 billion with up to $25 billion more committed, and has Anthropic to thank for more than half of its record profits last quarter. Media reports say that Amazon researchers tested Fable once it was released, found they could get Fable to cough up software vulnerability information by just rephrasing the question, carried it to the White House, and the export control followed.

Why would a company tattle-tale on a product it owns a slice of? The benefit-of-the-doubt view is that they were genuinely unnerved by what they saw in Fable, and wanted to keep the U.S. government informed. Some reports suggest that they brought the information to Anthropic before the White House.
But other analysts are more cynical: after all, Amazon owns a slice of the competition, too. The Fortune 1 company sells rival models through its Bedrock platform and is running the same race Anthropic is; so Amazon “will view Anthropic as the enemy in many cases.” 

Dion Hinchcliffe, an analyst who tracks enterprise AI at the Futurum Group, added that Amazon is losing the frontier lab race, and so it doesn’t much hurt them if the clear leader of the pack trips up. 

How dangerous the flaw even was is itself in contestation. Anthropic called the vulnerabilities minor and said rival models—OpenAI’s GPT-5.5 among them—could surface the same thing with no jailbreak at all. Dozens of security executives signed a letter Sunday, organized by former Facebook security chief Alex Stamos, calling the capability a normal feature of any model built to write secure code. Since the models the government flagged as dangerous, said Veracode co-founder Chris Wysopal, who signed the letter, are the same ones companies use to find flaws in their own code, taking it offline cuts both ways, he said: hurt the attackers and you hurt the defenders too.

That doesn’t mean you give up on finding bugs, but jailbreaks, Wysopal said, are a permanent cat-and-mouse game that no useful model ever fully wins; the normal fix is to tell the company and let it patch, not to invoke export controls. “There’s nothing open and public about how we are determining this,” he said. 

This story was originally featured on Fortune.com