Anthropic removed access to its freshly launched AI models Mythos 5 and Fable 5 over the weekend following sweeping restrictions imposed by the U.S. government.
In a statement, the lab said federal officials informed it on Friday afternoon that they had discovered a method for “jailbreaking” Fable 5, allowing users to bypass the safeguards Anthropic built to limit misuse. When Mythos was first announced, Anthropic limited distribution to a handful of government agencies and tech professionals because of the model’s capacity to uncover cybersecurity flaws.
The government applied export‑control measures to the products, which Anthropic said obliges it to block any foreign nationals—whether inside or outside the United States—from accessing the two models. The only way to enforce this, the company explained, was to shut the models down completely.
Anthropic later said it disagreed that the demonstrated jailbreak justified such a broad response and warned that a precedent like this could “essentially halt” the rollout of future frontier AI models if applied widely.
“As we have stated publicly, we believe the government should have the ability to block unsafe deployments, as part of a statutory process that is transparent, fair, clear and grounded in technical facts,” Anthropic wrote. “This action does not adhere to those principles.”
A spokesperson for Anthropic declined further comment beyond the statement. Officials from the U.S. Department of Commerce did not immediately respond to requests for comment.
Reports indicate Anthropic sent staff to Washington to negotiate with the Trump administration. The Wall Street Journal said Anthropic executives spent several hours on calls Saturday with Secretary of Commerce Howard Lutnick and National Cyber Director Sean Cairncross.
AI safety and jailbreaks
This is not Anthropic’s first clash with Washington over its models. Earlier this year, the Defense Department labeled Anthropic a “supply chain risk” after the company refused to permit its Claude models for fully autonomous weapons or mass domestic surveillance, while the Pentagon sought unrestricted use.
David Sacks, co‑chair of the President’s Council of Advisors on Science and Technology and former “AI and crypto czar” of the Trump administration, dismissed any link between the current dispute and the Pentagon issue. In a lengthy X post, Sacks said the administration believes the matter “should be easily resolved” and that “the ball is in Anthropic’s court.”
Sacks added that the government does not accept Anthropic’s view that the jailbreak is minor, noting the company’s reputation for AI safety. “In this case, Anthropic prioritized the continued offering of the consumer model over safety,” he wrote.
Ayham Boucher, executive director of AI strategy and innovation at Cornell University, said the problem is not simple. “All models can be jailbroken,” he explained, noting that a jailbreak lets users circumvent a model’s security controls by prompting it to perform disallowed actions—often via role‑play scenarios. For example, a model might not give direct instructions for robbing a bank, but could describe the process if asked within the context of a screenplay.
As models become more sophisticated, developers have bolstered defenses, yet the battle is ongoing. Anthropic acknowledged this with Fable 5, requiring users to allow the company to retain their interactions for 30 days—a practice that can deter some businesses but also provides data to detect and patch jailbreak attempts.
“They decided they can monitor, and then they can learn and make it a little harder,” Boucher said. “It’s not going to be perfect.” He warned that merely asking Anthropic to patch the hole reflects a misunderstanding of the technology, because “there will be another hole.”
No roadblocks to reaching Mythos‑level capability
While Anthropic is at the center of this episode, other labs are developing similarly capable models for coding and cybersecurity. Some researchers argue Mythos is not dramatically better at finding flaws than, for instance, OpenAI’s GPT‑5.5.
Models excel at cybersecurity tasks because coding and mathematics scale more easily than creative writing, which is harder to verify objectively. This trend applies across the industry, Boucher noted, and competitors—including OpenAI, Google, China’s DeepSeek and Alibaba—are quickly closing the gap.
The Trump administration has emphasized the need for American firms to stay ahead of China, but the sudden suspension of Fable and Mythos could hurt U.S. competitiveness. Export controls not only prevent foreign customers from using the models but also restrict foreign‑national employees at Anthropic, who are essential to future development.
“I can’t see how this is not slowing down model development,” Boucher said.
Regulatory backdrop
Washington’s stance appears to be shifting toward tighter oversight of advanced AI, though the approach remains fragmented. Last year, the administration released an AI action plan that framed limited regulation as key to maintaining “global AI dominance.” The Mythos incident seemed to catch the White House off guard.
Earlier this month, President Donald Trump signed an executive order—voluntary in nature—asking developers to submit frontier models for federal review before release if they posed security risks. Samir Jain, vice president of policy at the Center for Democracy and Technology, observed that the voluntary window was short‑lived.
Jain stressed that cybersecurity concerns are legitimate and that the government has a stake in protecting critical systems. However, he criticized Friday’s move as arbitrary and opaque, noting the lack of a clear legal basis or transparent process.
First‑Amendment considerations also arise, Jain added. AI models generate speech, and developers make editorial choices, so any regulation must respect freedom of expression for both the creators and the users.
“That doesn’t mean the government can’t regulate these models,” Jain said. “But it does put even more premium on making sure whatever regulation is done is done in accord with the rule of law.”
