The US government has instructed Anthropic to cease operations of Fable 5 and Mythos 5.

      **Summary:** On June 12, the US government directed Anthropic to halt access to its top-tier AI models, Fable 5 and Mythos 5, due to national security worries stemming from a reported jailbreak. Although Anthropic complied, it publicly criticized the decision as excessive and cautioned that if this action became standard across the industry, it would cease all deployments of advanced models. This unprecedented shutdown occurred just three days after the launch of Fable 5.

      The US government mandated that Anthropic disable access to its two most advanced AI models, Fable 5 and Mythos 5, marking a potential first for Washington in taking a commercial AI product offline. The order, which Anthropic claims it received at 5:21 PM ET on June 12, cites national security concerns and requires the firm to cut access for all foreign nationals, including its own employees. Since Anthropic cannot efficiently differentiate between foreign and domestic users in real time, this effectively means a complete shutoff of both models for all customers globally. Access to Anthropic's other models remains unaffected.

      From launch to shutdown in three days

      Anthropic introduced Fable 5 on June 9 as its first “Mythos-class” model available to the public. Fable 5 serves as the public version, employing safety classifiers that direct flagged inquiries to the less powerful Claude Opus 4.8. In contrast, Mythos 5, the equivalent model with cyber protections removed, was restricted to authorized cybersecurity professionals and critical infrastructure operators.

      The latest insights from the EU tech sector, a tale from our esteemed founder Boris, and some questionable AI art. It’s free, delivered weekly to your inbox. Sign up now! Both models were already under scrutiny from the Trump administration, which had previously sought to hinder Anthropic’s broader Mythos rollout. The company offered Fable 5 for free to Pro, Max, Team, and Enterprise subscribers until June 22.

      The jailbreak controversy

      Reportedly, the Commerce Department took action after an unnamed company alleged it discovered a method to jailbreak Mythos, according to Axios. Anthropic contends it has only received verbal claims of a narrow, non-general jailbreak that involves prompting the model to interpret a specific codebase and correct software errors.

      Anthropic evaluated the method and stated that the vulnerabilities identified were minor and already publicly acknowledged. The company also asserts that other models, including OpenAI’s GPT-5.5, can uncover the same flaws without any bypass, although TNW could not independently confirm this comparison.

      The firm emphasizes the difference between narrow jailbreaks, which can retrieve specific information in certain conditions, and universal jailbreaks, which broadly undermine a model’s safeguards. Anthropic claims that no tester has discovered a universal jailbreak for Fable 5, although the UK AI Safety Institute did manage to create a partial jailbreak for single-turn queries within mere hours of testing.

      Extensive red-teaming efforts beforehand

      Prior to the launch, Anthropic reveals it collaborated with the US government, the UK AISI, and several private organizations to conduct extensive red-teaming on Fable’s safeguards for thousands of hours. An external bug bounty allegedly yielded no universal jailbreaks over more than 1,000 hours of testing.

      The company also implemented a controversial 30-day mandatory data retention policy for all traffic related to Fable and Mythos, overriding previous zero-retention agreements with clients. Anthropic argues that this retention is essential to recognize patterns of misuse that only emerge across multiple requests.

      Anthropic responds

      In a notably assertive public statement, Anthropic expressed disagreement with the notion that a narrow potential jailbreak warrants “recalling a commercial model deployed to hundreds of millions of users.” The firm cautioned that applying this standard industry-wide would “effectively halt all new model deployments for all frontier model providers.”

      The company called for any governmental authority to block unsafe deployments to be based on “a statutory process that is transparent, fair, clear, and grounded in technical facts,” asserting that this directive does not meet that standard.

      Rising tensions

      This directive occurs amidst ongoing tension between Anthropic and Washington. Earlier this year, the Pentagon classified Anthropic as a national security supply chain threat, even as the NSA continued utilizing Claude due to a lack of alternatives. Anthropic is currently suing the government regarding that classification. The same administration that labeled the company a security threat has also encouraged banks to adopt its technology and authorized the NSA to continue using Mythos on classified networks.

      This contradiction highlights the government’s internal conflict regarding how to manage a company whose AI capabilities it simultaneously fears and relies upon.

      Implications for the industry

      If this precedent is upheld, any frontier AI model could face a similar recall due to a single reported jailbreak, regardless of its severity. Such a possibility is likely to instill apprehension within an industry that has spent the past year rapidly deploying increasingly advanced models amid escalating government scrutiny.

      Anthropic maintains that it believes the directive is a misunderstanding