Anthropic on Wednesday unfroze access to its advanced AI models Fable and Mythos, following the Trump administration's lifting of related export control restrictions.

The return of the two models—with Fable 5 open to general users, while the more powerful Mythos 5 remains restricted to a trusted partner consortium as before the ban—marks significant progress in ongoing negotiations between the Trump administration and AI companies over the responsible deployment of frontier models. In recent months, the U.S. government and the tech industry have been at odds on this issue.

On Tuesdayin a statement announcing the resolution of one of the most intense conflicts in the AI field, Anthropic insisted that its models have always been safe and argued that the government exaggerated the severity of the situation—hinting that tense dialogues will continue in the future over how to balance providing advanced capabilities to defenders while preventing U.S. adversaries from obtaining them.

The U.S. Department of Commerce issued the export control ban after Amazon warned the government that Fable's safety measures could potentially be bypassed. On Tuesday, Anthropic reiterated its argument—a leading coalition of cybersecurity expertsalso made the same point—that less capable AI models have the same issues, but the company also said it has "quickly taken action to address the reported bypass issues."

Anthropic said that after working closely over the past two weeks with "the government and other partners, including Amazon," the company "trained an improved safety classifier to identify and block the behaviors described in the report."

Anthropic added that researchers at the U.S. National Institute of Standards and Technology's AI Standards and Innovation Center "have tested both our previous and new safety measures and unanimously found them to be very robust."

Meanwhile, the company warned that these changes could have some negative effects on cybersecurity researchers seeking assistance with defensive work.

"The new classifier... comes at the cost of more frequently flagging benign requests in routine coding and debugging tasks," Anthropic said. "As with all our safety measures, we will continue to refine it to better distinguish genuine abuse from legitimate requests and reduce false positives."

Calls for a more formal review process

Although the controversy surrounding Fable and Mythos may have ended, the AI industry remains concerned about the Trump administration's ad hoc approach to reviewing the availability of frontier models.。

Trump recentlyissued an executive orderestablishing a process for frontier AI companies to provide the government with early access to certain particularly powerful models. On Wednesday, Anthropic said it would provide such early access for "models that substantially advance the capability frontier in national security-related areas." The company also said it would share intelligence on how hackers abuse its tools and participate in the vulnerability information-sharing center established under Trump's directive.

Anthropic emphasized its commitment to working closely with the government to address potential AI safety risks. The company said it is "significantly expanding" collaboration with federal agencies, including dedicating specialized personnel and computing resources. It also pledged to "work with government and industry peers to jointly develop a voluntary safety and evaluation standard for frontier model providers."

Anthropic also noted that there is currently a "lack of recognized standards" for classifying the severity of jailbreaks—an important prerequisite for any formal model review process.

"A common standard for evaluating AI jailbreaks would help us and other companies safely release new models while also allowing users to fully leverage their advanced capabilities," Anthropic said.

To that end, the company announced it is working with Amazon, Google, Microsoft, and other members of Project Glasswing (through which Anthropic grants Mythos access to vetted organizations) to develop a "consensus framework" for jailbreak classification and response. The company said the framework envisions rating each potential jailbreak on four criteria, including the ease of discovering the bypass method and the extent of additional model capabilities unlocked by the bypass.