On June 12, Anthropic announced that it had received directives from the U.S. government to suspend access to its newly launched models, Fable 5 and Mythos 5, for “foreign nationals, whether in the U.S. or elsewhere, including Anthropic employees.”
The AI startup stated that the U.S. government cited “national security” as the reason but did not provide further details. “We believe that the U.S. government should have the ability to prevent the deployment of dangerous applications based on a transparent, fair, clear, and technically grounded process. This action does not meet those principles,” the company’s statement read.
On June 14, David Sacks, co-chair of the U.S. President’s Council of Advisors on Science and Technology, revealed that a trusted partner of the U.S. government and Anthropic had tested Fable and discovered a jailbreak method to bypass the safety barriers that separate this commercial model from the unrestricted cybersecurity capabilities of Mythos.
The U.S. government requested that Anthropic patch the vulnerabilities or withdraw Fable 5 until the issues were resolved, but CEO Dario Amodei reportedly refused. “The administration then had to reluctantly issue export control directives. They are willing to lift the restrictions as soon as the issues are resolved. The decision now rests with Anthropic,” he stated.
Logo of the Claude application. Photo: Luu Quy
Sacks criticized Anthropic’s actions as inconsistent with the position of a lab that prioritizes safety and has called for the regulation of Mythos as a cybersecurity weapon. He indicated that the decision to impose restrictions was not related to previous tensions between Anthropic and the government, adding that Washington has always valued the technology of this AI startup and considers the issue “easily resolvable.”
Anthropic has not commented on this information.
The WSJ reported, citing anonymous sources familiar with the matter, that the control measures were partially triggered by cybersecurity research documents from Amazon, as well as discussions between Amazon CEO Andy Jassy and the White House. Amazon had previously invested billions of dollars in Anthropic.
According to these documents, a person could manipulate Fable 5 to provide information used for cyberattacks with just a few commands. Shortly after CEO Jassy shared this discovery with the White House, the U.S. government decided to block foreign nationals from using Mythos 5 and Fable 5. The situation was further complicated by the fact that many Anthropic researchers were born abroad, preventing them from accessing their own products.
Semafor quoted a source close to the White House asserting that Amazon had alerted the White House about the jailbreak method through CEO Jassy. The U.S. government acted swiftly, partly due to concerns that a group of Chinese hackers had accessed Mythos, leading to the risk of the model being reverse-engineered and replicated.
Logo of Claude Fable 5. Photo: Anthropic
Amazon did not confirm the information but stated that the U.S. government frequently consults them about security threats, emphasizing that the company does not comment on discussions with the White House.
Anthropic expressed skepticism regarding the assertion that the issue lay with the “jailbroken model.” They claimed that similar vulnerabilities could easily be discovered through existing models in the market, including GPT 5.5.
Some security researchers support this hypothesis. Katie Moussouris, founder and CEO of cybersecurity firm LutaSecurity, stated that she reviewed Amazon’s documents and confirmed, “This is not a jailbreak.” Kate Koren, a former U.S. Department of Commerce official, assessed that the decision to impose restrictions stemmed partly from the White House’s dissatisfaction with Anthropic.
The ban came just days after Anthropic announced Fable 5 and Mythos 5, two models the company touted as the most advanced by many metrics in the AI industry. They are based on the Claude Mythos Preview, which had previously raised concerns among U.S. officials due to its advanced cybersecurity capabilities.
This is the latest conflict between Anthropic and the U.S. government following a dispute with the Pentagon earlier this year. The U.S. Department of Defense then declared Anthropic a risk to the supply chain, meaning the company was viewed as a national security threat. This designation is typically reserved for foreign adversaries, requiring defense contractors to certify that they do not use the Claude model in contracts with the military. Anthropic has sued the Trump administration to reverse this decision, and the litigation process is still ongoing.
Diep Anh compiled the report.
