The federal government has shut down Anthropic’s latest model, Claude Fable 5, and Claude Mythos 5 AI models, citing “national security” concerns and its potential to be exploited by bad foreign actors. 

The shutdown occurred Friday, immediately after the release of the models, after Amazon warned a day earlier that researchers had identified ways to “jailbreak” the models, prompting AI to override its own safety regulations. 

The “export control directive,” a legal order the Commerce Department imposed to restrict the technology from leaving the country, would have prevented non-US nationals from accessing the models. This reflects a critical point as the government grapples with how far it will go to regulate frontier AI models.

In a social media post on Friday, Anthropic said that “the net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.”

Anthropic had assured that it had safeguards in place to keep the models from being exploited for cyberattacks, and notified the government many times about the impending release of the models. However, top administration officials began to doubt how secure the model really was after Amazon issued its warning.

On Friday, administration officials reportedly spent hours Friday asking Anthropic to voluntarily shut down access to Fable 5 and Mythos 5 models. According to Axios, an official said that these efforts were unsuccessful.

That afternoon, Anthropic received a call from administration officials around 1 p.m., saying that the government would move forward with export restrictions unless the company acted, giving them roughly 90 minutes to do so.

At 5: 21 p.m. ET, the Commerce Department’s enforcement order restricted access to the models by foreign nationals, including U.S. allies and foreign nationals in the U.S. Anthropic then disabled access to its models globally, as it could not immediately separate eligible from ineligible users. 

The export control has ultimately forced Anthropic to withdraw its new AI models three days after its release. While government officials were quick to praise Trump’s export controls, the cyber world has pushed back.

According to Axios, CEO Dario Amodei and other Anthropic officials reportedly spoke with the administration after the government imposed export rules, where they stressed that the Amazon jailbreak was relatively simple and could be achieved by other models. 

“These vulnerabilities all appear relatively simple, and we have found that other publicly available models are able to discover them as well without requiring a bypass,” Anthropic said on Friday in a blog post. They also said in the statement that the company disagreed that “the finding of a narrow potential jailbreak should be cause for recalling a commercial model.”

Cynthia Kaiser, former deputy assistant director of the Federal Bureau of Investigation’s Cyber Division, wrote on LinkedIn Sunday, “The only way to defend against bad actors gaining AI capabilities is to use those same capabilities to stop them.” 

The restrictions also point to a pressing dilemma in the AI race, the urgent need to secure safeguards around frontier models before China catches up. 

Recent estimates suggest that the U.S. has six to 12 months before Beijing also gained access to a frontier model comparable to Mythos that could potentially be used as a cyberweapon, according to Politico.

In April, Anthropic had said that Mythos, its most advanced model at the time, was too good at hacking for immediate release, showing unusual skill at finding software bugs and exploiting them to take control of systems. Instead of a public launch, Mythos was restricted to a vetted group of roughly 150 organizations around the world, including critical infrastructure operators, cybersecurity researchers and government agencies.

Fable is the same basic model, but implements further safeguards to prevent it from being used for cyberattacks. Before its launch, Anthropic said its team had spent thousands of hours trying to get Fable to release information on banned topics, but had not identified universal jailbreaks, which bypasses an AI model’s safety guardrails across the board.

The Trump administration has been relatively hands-off with regulating AI to avoid suppressing innovation and to stay competitive with China. 

Yet on June 2, Trump signed an executive order that encourages AI companies to submit new models for voluntary review at least 30 days before public release, following Anthropic's warning that the “fallout—for economies, public safety, and national security” from new AI models “could be severe."