Anthropic has suspended access to its latest AI models, Fable 5 and Mythos 5, following a directive from the U.S. Commerce Department that invokes national security export controls. This order prohibits the distribution of these models not only to individuals outside the United States but also to any foreign nationals within the country, including Anthropic’s non-citizen employees.
In light of the directive, Anthropic announced it had no option but to disable these advanced models for all users, while its less powerful Claude series, including the recent Claude Opus 4.8, remains accessible. The company expressed regret over the interruption to its customers, stating that it views this situation as a misunderstanding and is actively working towards restoring access.
According to a blog post by Anthropic, the directive was received at 5:21 PM Eastern Time, lacking detailed information on the national security concerns raised by the government. However, the company noted that officials indicated the decision stemmed from knowledge of a method that could circumvent Fable 5’s protective measures, which were intended to restrict access to Mythos’s advanced cybersecurity functions.
Anthropic maintains that the alleged jailbreak referenced by the government was limited to a specific situation and does not undermine the overall safeguards of Fable 5. They assert that similar vulnerabilities could exist in other publicly accessible models, such as OpenAI’s GPT-5.5, which are not bound by these national security restrictions. Anthropic contends that the narrow potential jailbreak should not warrant the recall of a commercial model widely utilized globally and argues that applying such a standard could hinder the deployment of new models across the industry.
The AI sector and policy analysts have expressed astonishment at this extraordinary move by the U.S. government. Some commentators suggest it might be a continuation of punitive actions originating from the Trump administration against Anthropic. Earlier this year, then-President Trump instructed all federal agencies to halt the use of Anthropic’s models after the company declined to comply with certain contract terms favored by the Pentagon. These terms imposed a broader use of AI models, including for military purposes that Anthropic sought to avoid.
The Pentagon labeled Anthropic a “supply chain risk” in March, subsequently banning its models’ use within the military framework and restricting defense contractors from utilizing its technology in government contracts. Anthropic is currently contesting this designation in court.
As the company remains poised for a public listing, it faces potential investor skepticism due to the government’s recent actions, which could impact investor confidence regarding Anthropic’s ability to remain at the forefront of AI model development. Some tech policy figures associated with the Trump administration have openly criticized Anthropic, accusing it of adopting a “woke” ethos and exploiting regulatory mechanisms through fear tactics.
Experts in AI policy have varied opinions on the implications of the government’s actions. Some view them as overly extreme, questioning the rationale behind a national security approach that simultaneously suggests exporting advanced AI technologies to countries like China while imposing stringent restrictions at home.
Conversely, a faction of critics argues that Anthropic’s previous claims about the dangers of its models may have laid the groundwork for the government’s response, suggesting that companies must be cautious about how they characterize their technologies. Past instances of government restrictions on software exports, such as encryption technology, have typically failed, indicating a complex history surrounding regulations in this sector.
Further complicating the landscape, critics warn that these actions may drive international talent, particularly those with ties to China, away from American AI firms, raising concerns about the long-term viability of U.S. leadership in AI innovation. Some speculate that the hiatus in development, encouraged by those wary of AI’s potential risks, could result in a slowdown in progress that safety advocates might welcome.
