The U.S. government ordered Anthropic to immediately shut off access to two of its most powerful AI models, Claude Fable 5 and Claude Mythos 5.
The directive, which was received on Friday at 5:21 pm ET, cited national security concerns, forcing the company to disable both models for all users globally. The government’s action effectively overrides export control restrictions that had nominally aimed only at foreign nationals. Anthropic stated that access to its other AI models remains unaffected.
Mythos is described as Anthropic’s most capable model. It was previously kept tightly restricted because of its exceptional ability to identify security vulnerabilities in software. According to the company, Mythos identified flaws in every major operating system and web browser it tested. To manage this capability, Anthropic launched a controlled program called Project Glasswing, giving access only to approximately 50 vetted organizations, including Apple, Google, Amazon, Microsoft, and CrowdStrike, for defensive cybersecurity work.
Anthropic’s second major offering was Fable 5, released just three days prior. This version was designed as a safer commercial alternative to Mythos. It included guardrails that block responses in high-risk areas such as cybersecurity and biology, allowing it general release. Vals AI reported that upon its debut, Fable 5 was the most capable public AI model available.
While the government’s order is framed as an export control action restricting foreign access, Anthropic highlighted broader concerns. The underlying worry, according to the company’s blog post, related to a potential jailbreak of Fable 5—a narrow, non-universal vulnerability that might involve prompting the model to read specific codebases and identify software flaws. Anthropic pointed out that this level of capability is already widely available in other public models, including OpenAI’s GPT-5.5, and is routine for cybersecurity professionals.
Anthropic also emphasized its safety safeguards, noting they operate through independent classifier systems separate from the model itself, thereby maintaining underlying protections even if a user manages to bypass an initial refusal.
The company expressed significant disagreement with the government’s directive. Anthropic stated: “We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.” It argued that applying this standard industry-wide would essentially halt new model deployments for all frontier model providers.
Observers noted an irony, as the caution Anthropic displayed by restricting Mythos—a model it promoted as dangerously powerful—had attracted exactly the kind of government scrutiny that could disrupt its business. Meanwhile, OpenAI’s Sam Altman had previously criticized Anthropic’s handling of Mythos in April, calling it “fear-based marketing.”
