The U.S. government has ordered Anthropic to halt overseas access to its advanced artificial intelligence models, Claude Fable 5 and Claude Mythos 5, citing national security concerns.
The restrictions apply to all non-U.S. citizens worldwide, including Anthropic's own foreign employees. Anthropic immediately criticized the decision, describing it as arbitrary and lacking clear principles.
On June 12 (local time), Anthropic announced that U.S. national security authorities had issued export-control directives and that, in compliance with the order, it would temporarily suspend access to Fable 5 and Mythos 5 for all customers. Other models, including Claude Opus 4.8, remain available.
The affected models include Fable 5, a safety-focused version designed to enable broader public access to Mythos-class capabilities, and Mythos 5, a security-oriented model.
Fable 5 incorporates safeguards against high-risk misuse. It restricts queries related to cybersecurity attacks, biological and chemical threats, and suspected attempts to extract capabilities from competing AI systems through model distillation. Despite these protections, the U.S. government issued its control directive at 5:21 p.m. Eastern Time.
Anthropic stated that the government's notice did not specify any concrete national security concerns. Based on its own assessment, the company believes authorities may have obtained a method for bypassing or “jailbreaking” Fable 5.
According to Anthropic, a review of demonstrations using a particular technique uncovered several previously known minor vulnerabilities. The company argued that these issues are relatively easy to address and that similar weaknesses can be identified in other publicly available AI models.
Anthropic emphasized that Fable 5 includes robust security measures designed to significantly reduce the risk of misuse in cybersecurity-related activities. The company noted that some users had even complained that the safeguards were overly restrictive.
Before launching Fable 5, Anthropic said it collaborated with the U.S. government, the UK AI Safety Institute (AISI), external organizations, and internal teams to conduct thousands of hours of red-team testing to validate the model's security protections.
The company argued that no AI developer can completely eliminate the possibility of jailbreaks. It said all industry-standard safety systems remain vulnerable under certain circumstances and that broadly applicable jailbreak techniques may eventually emerge for any advanced AI model.
To minimize such risks, Fable 5 employs a defense-in-depth strategy. Even if a jailbreak were possible, Anthropic said the model was designed to make exploitation costly and difficult, while extensive monitoring measures--including retaining customer data for 30 days--allow rapid detection and mitigation of malicious activity.
Anthropic said the model was released only after extensive preventive measures had been implemented and that it has not received reports of any widespread or severe jailbreak vulnerabilities. As a result, the company suggested that the U.S. government's decision is difficult to justify.
The company added that its analysis of the suspected jailbreak method indicates that other frontier AI systems with comparable capabilities--including OpenAI's GPT-5.5--could potentially be vulnerable to similar techniques.
Anthropic said it would provide additional details within 24 hours, criticizing the government's action as inconsistent with established principles.
“The government has the authority to block unsafe deployments through a transparent, fair, and clearly defined legal process grounded in technical facts,” the company said. “However, this action does not meet those standards.”
Anthropic added that it believes the directive may have resulted from a misunderstanding, but that it had nevertheless blocked access to Fable 5 and Mythos 5 for all users in order to comply with the government's legal order. The company said it is working to restore access as quickly as possible.
The company also argued that recalling a commercial AI model already distributed to hundreds of millions of users would be unjustified based solely on the discovery of an extremely limited jailbreak vulnerability. Anthropic warned that if such a standard were applied across the industry, the release of virtually all new AI models could be halted.
Anthropic has previously clashed with the U.S. Department of Defense over the military use of AI. The dispute arose after the department reportedly sought broader authorization for applications such as large-scale AI-powered surveillance and fully autonomous weapons operating without meaningful human control--requests Anthropic said it could not accept.
At the time, Anthropic CEO Dario Amodei publicly stated that such proposals were “something I simply cannot support in good conscience.”
Following the dispute, the U.S. Department of Defense designated Anthropic as a supply-chain risk company. Anthropic subsequently filed a lawsuit seeking to overturn the designation, naming the Department of Defense, the U.S. General Services Administration (GSA), and Defense Secretary Pete Hegseth among the defendants.
