The Government Just Forced Anthropic's New AI Model Offline

In Brief

Anthropic released Claude Fable 5, its new flagship model, then disabled it for everyone days later after the US government issued national-security export controls barring its distribution to foreign nationals. The trigger appears to be a jailbreak that researchers at Amazon used to surface dangerous cyber capabilities, and the episode is the first time Washington has yanked a just-released frontier model offline.

What Happened

The US government forced Anthropic to take its newest flagship AI model offline days after launch. Anthropic released Claude Fable 5 as a major step up in capability, a model built from its more powerful Mythos class but made safe for general use. Then the government issued national-security export controls barring Anthropic from distributing Fable 5 and Mythos to foreign nationals, meaning people outside the US and non-citizens working inside it, a group that would include some of Anthropic's own employees.

Because Anthropic cannot selectively filter who uses a model based on citizenship, it disabled Fable 5 and Mythos for all users. The launch had been impressive on paper. Anthropic claimed state-of-the-art results in coding, vision, and long-context work, and Stripe said it used Fable 5 to compress months of engineering into days, completing a 50-million-line code migration in a single day.

The intervention appears to trace back to Amazon. According to Axios, Amazon researchers, part of a group testing the models, found a way to jailbreak Fable 5 (a jailbreak is a prompt trick that bypasses a model's safety guardrails) and access portions of the more powerful Mythos capabilities that can aid cyberattacks. The government then gave Anthropic 90 minutes to take the models down. Anthropic disputes the framing, calling the vulnerabilities minor, already replicable with other public models, and not a universal jailbreak. SmarterX founder and CEO Paul Roetzer broke down what the move means for businesses on Episode 219 of The Artificial Intelligence Show.

The Key Numbers

90 minutes - Time the government gave Anthropic to take Fable 5 and Mythos offline

50 million - Lines of code Stripe migrated in a single day using Fable 5

3 to 6 months - Roetzer's estimate for China matching Mythos-level capabilities

$25 billion - Amazon's pledged investment in Anthropic on top of $8 billion already in

~$1 trillion - Anthropic's approximate current valuation

Why Businesses Now Have to Worry About Model Control

The model you use is not the model the labs have. The version of any frontier model available to the public is a guardrailed, safer cut of something far more capable sitting inside the lab. Fable 5 is the same underlying model as Mythos, with prompting tricks able to unlock the more dangerous capabilities.

"The models that you and I use every day are not the versions that they have in the labs," says Roetzer. "They have more powerful versions, more generally capable versions of these models that get guardrails on them so they're safe for us to use."

Shutting one model down does not contain the capability. Whatever made Mythos dangerous enough to pull will not stay contained to Anthropic.

"Whatever Mythos is, whatever its capabilities are that caused it to be taken down by the government, China is going to have those capabilities within three to six months, if they don't already," Roetzer says. "An open source version of Mythos will likely come from someone within six to 12 months."

His read is that anything bad enough to force a shutdown today becomes universally available within the year, and the government is not prepared for that.

The mixed messaging undercuts the official story. White House AI czar David Sacks posted that anyone who thinks the administration is targeting Anthropic does not understand the situation, and that the government values the company.

That is a hard message to square with Secretary of War Pete Hegseth, who posted that, three months ago, he had "kicked Anthropic AI out of our building forever" and that "every passing day proves why that was the right move." As Roetzer puts it, "those two things can't both be true."

The bigger risk is for the companies building on these models. The government has no formal AI regulations in place, yet it is using export controls and supply-chain risk designations to decide what everyone can access. That lands directly on businesses that have woven AI into their workflows.

"If we're relying on Anthropic or OpenAI or Google, and we infuse our workflows, our AI's infused into workflows in our jobs, and then the government decides all of a sudden, 'Oh wait, that model's not safe, gone.' And now you're like, 'What the hell?' We just built our business around this whole thing."

— Paul Roetzer, founder and CEO of SmarterX, Episode 219 of The Artificial Intelligence Show

SmarterX Take

There is an irony underneath all of this. One day after the Fable 5 release, Anthropic CEO Dario Amodei published an essay arguing that frontier models, like airplanes, should be required to pass technical testing and that the government should be able to block or reverse a release that presents unacceptable risks. Two days later, the government did almost exactly that to his own company. Anthropic has warned that if this standard gets applied across the industry, it would effectively halt all new model deployments from every frontier provider.

For business leaders, the lesson is about concentration risk. When the model powering your core workflows can vanish on 90 minutes' notice over a national-security call, single-vendor dependence becomes a strategic liability. Roetzer's conclusion is that more companies will need to take seriously the idea of building on and controlling their own models, a shift that could swing the market back toward open source. We are now in an era where the government recognizes the value of these models, wants full access to them, and wants to dictate when something is not safe enough to ship.

What to Watch

Uncertainty is the real story for the markets. Anthropic's technical staff traveled to Washington to meet with White House officials, and the administration says access should return once the vulnerability is patched. But the precedent now exists, and it creates a fog of unknowns around the future of model deployment, including the planned IPOs of Anthropic and OpenAI. As Roetzer notes, Wall Street hates uncertainty, and there is now plenty of it.

Access is also starting to split along who can pay. Even before the government stepped in, Fable 5 was drawing scrutiny because it sits outside Anthropic's standard plans as pay-as-you-go usage after a short preview period. Put the pricing and the politics together and the trajectory points toward a world where the most capable models are gated by both money and government approval, a combination business leaders should plan around now.

How Ready Is Your Organization to Govern AI?

Only 13% of organizations have the governance foundations to scale AI responsibly, according to the 2026 State of AI for Business Report. That means an AI council, an AI roadmap, generative AI policies, and an AI ethics policy all in place. A third of organizations have none of them. When the rules can change on 90 minutes' notice from Washington, the companies without that foundation are the ones most exposed to sudden disruption.

The 2026 State of AI for Business Report is built on more than 2,100 responses across roles, functions, and industries, and it maps exactly where organizations stand on readiness and governance. If a single government decision could pull a model out from under your workflows tomorrow, the report is the clearest benchmark for whether your organization is built to absorb that kind of shock. Read the full report →