US tells Anthropic to suspend Fable 5, Mythos 5 access for foreign nationals

In a significant move, the U.S. government has instructed Anthropic to suspend access to its Fable 5 and Mythos 5 AI models for all foreign nationals, including the company’s own employees who are not U.S. citizens. This directive prompted Anthropic to disable these models for all users to ensure compliance. Anthropic revealed that this directive fell under national security regulations and affects foreign nationals regardless of their location. The company reassured its customers that access to its other AI models remains unaffected. "To comply with the order, we must disable Fable 5 and Mythos 5 for everyone," Anthropic stated. This intervention by the U.S. government is notable, especially coming just days after Anthropic launched its Mythos-class systems. Fable 5 was designed for public use, while Mythos 5 had stricter access controls. Anthropic received the government’s order on June 12 at 5:21 PM ET but mentioned that the letter did not clarify the specific national security concerns. According to Anthropic, the government expressed concern about a potential method to bypass or 'jailbreak' Fable 5. The company indicated that it had seen a demonstration of this technique, which only exploited a few previously known minor vulnerabilities. Anthropic contended that these vulnerabilities were relatively simple and could also be discovered using models available to the public. The company emphasized that it had implemented robust safeguards for Fable 5, particularly to prevent cybersecurity misuse. They collaborated with the U.S. government, the UK AI Safety Institute, private organizations, and internal teams to rigorously test the model before its release. "Our tests demonstrated that Fable's protections are significantly more effective than those of any models previously deployed," Anthropic claimed. Despite these assurances, Anthropic acknowledged that no testing had uncovered a 'universal jailbreak' that could compromise the model's security across the board. They suggested that achieving perfect jailbreak resistance is currently unattainable for any AI provider, and all industry safeguards have vulnerabilities to non-universal jailbreaks. In response to the government’s directive, Anthropic reiterated its commitment to a 'defense in depth' strategy aimed at making jailbreaks either limited in scope or costly to implement. This approach also includes monitoring to detect and halt any potential attacks, which is why the company mandates a 30-day retention of customer data for Fable, despite the associated costs. Anthropic disagreed with the government’s reasoning, asserting that it had not been presented with concrete evidence of a serious jailbreak leading to harmful outcomes. They noted that the government had only provided verbal indications of a narrow, potential jailbreak that involved merely asking the model to analyze specific code and rectify software issues. Anthropic also evaluated a report that they believed underpinned the government's directive, concluding that the demonstrated ability was already possible with other models, including OpenAI's GPT-5.5. While complying with the legal directive, Anthropic expressed dissatisfaction with the decision, arguing that the identification of a narrow potential jailbreak should not necessitate the recall of a commercial model used by millions. The company warned that if this standard were applied across the industry, it could effectively halt new model deployments altogether. Anthropic supports the government’s role in preventing unsafe AI applications but insists that such actions should follow a statutory process that is transparent, fair, and based on technical facts. They remarked, "This action does not adhere to those principles." The suspension of access to Fable 5 and Mythos 5 could disrupt users who have started to integrate these models into their operations, particularly in software development, research, and business automation. Anthropic is currently working on restoring access.