On June 13, 2026, Anthropic announced the immediate suspension of its advanced AI models, Claude Fable 5 and Mythos 5, for all users. This action follows a directive from the U.S. government to restrict access to these models for foreign nationals, both within and outside the United States, due to national security concerns.
The company received the government order at 5:21 p.m. ET, instructing the cessation of access to the specified models by foreign nationals. Anthropic expressed its belief that this directive stems from a misunderstanding and is actively working to restore access promptly. Notably, access to other Anthropic models remains unaffected by this export control measure.
Anthropic indicated that the government’s concerns may be linked to a perceived method of bypassing, or ‘jailbreaking,’ Fable 5. Upon review, the company observed that this technique was used to identify a limited number of previously known, minor vulnerabilities. These vulnerabilities appear relatively straightforward, and similar findings have been achievable using other publicly available models without necessitating any bypass methods.
This development comes shortly after the launch of Claude Fable 5 and its counterpart, Mythos 5. While both models share the same underlying architecture, Mythos 5 has certain safeguards lifted, particularly in cybersecurity domains. Mythos 5 is described as possessing the strongest cybersecurity capabilities among existing models and is currently accessible exclusively to a vetted group of cyber defenders and critical infrastructure operators.
Anthropic has emphasized the implementation of robust guardrails to prevent the misuse of its models for cybersecurity-related tasks. These safeguards include a set of safety classifiers designed to detect potential misuse, such as jailbreak attempts, and to prevent the main model from responding to such queries. The cybersecurity classifier, in particular, is engineered to block harmful requests related to cyber attack planning, exploit development, or defense evasion.
Recent findings from Anthropic’s Red Team highlight the capabilities of Mythos-class models in rapidly weaponizing newly disclosed software vulnerabilities. These models can transform N-day vulnerabilities into working exploits within hours or even minutes, significantly reducing the time frame from weeks to mere hours. This advancement suggests that frontier models are increasingly proficient at swiftly exploiting publicly disclosed flaws.
In response to the government’s directive, Anthropic has reiterated that no universal jailbreak methods have been developed against its latest models to date. Both internal and third-party red-teaming exercises have demonstrated that the company’s safeguards are substantially more effective than those of any previously deployed model. However, Anthropic acknowledges that achieving perfect jailbreak resistance is unattainable for any model provider, as all existing safeguards are susceptible to non-universal jailbreaks that are effective in very limited contexts or require additional effort to be exploited.
This situation underscores the delicate balance between advancing AI capabilities and ensuring national security. As AI models become more sophisticated, the potential for misuse increases, prompting governments to implement stringent controls. The current suspension of Anthropic’s models for foreign nationals highlights the need for ongoing dialogue between AI developers and regulatory bodies to navigate the complexities of AI deployment in a global context.
