In a significant move, the U.S. government has mandated Anthropic to immediately disable access to two of its advanced AI models, Claude Fable 5 and Claude Mythos 5, citing national security concerns. Anthropic confirmed on social media platform X that it has complied with the directive but expressed its belief that the government’s actions were misguided.
The order, which Anthropic received on Friday evening, compels the company to suspend access to these models for all users around the globe, instead of focusing only on foreign nationals as originally intended by the export control action. Notably, Anthropic’s other AI models remain unaffected by this directive.
The significance of this situation lies in Mythos, Anthropic’s most advanced AI model. First previewed in early April and subsequently kept under tight control, Mythos is reputed for its exceptional capabilities in uncovering security vulnerabilities in various software. Anthropic disclosed that Mythos has successfully identified flaws in every major operating system and web browser it tested. To ensure its responsible use, the company launched Project Glasswing, a controlled program allowing around 50 select organizations—including tech giants like Amazon, Apple, and Google—to utilize Mythos for defensive cybersecurity projects.
Just three days prior to the government’s order, Anthropic released Fable 5, which was designed as a safer, commercial version of Mythos. Equipped with guardrails to prevent high-risk outputs in areas such as cybersecurity and biology, Fable 5 quickly earned recognition as the most capable AI model available publicly, according to benchmark tests conducted by Vals AI.
Although the government presents its action as an export control measure aimed at restricting access for foreign nationals, Anthropic’s internal communications suggest that the core concern arises from allegations of a limited jailbreak of Fable 5. The company reported that the government has only provided verbal evidence of what they term a “potential narrow, non-universal jailbreak.” Anthropic contends that this jailbreak consists merely of prompting the model to analyze specific codebases and identify software flaws—capabilities that are not unique to Fable 5 and are already accessible in many widely used models, including OpenAI’s GPT-5.5.
Anthropic argues that its robust safeguards are facilitated through independent classification systems that operate separately from the model, ensuring that even if users manage to bypass certain restrictions, the foundational protective measures against hazardous outputs remain intact.
Despite these defenses, the government’s decision has sparked frustration within Anthropic. The company openly disagreed with the notion that the existence of a narrow jailbreak justifies recalling a commercially deployed model that serves millions. In their statement, Anthropic warned that extending this standard across the industry could effectively freeze the deployment of new models from all leading providers.
As Anthropic gears up for a potential initial public offering (IPO) this year, its image has been closely tied to its commitment to safety in AI development. Observers note the irony that the precautionary measures taken with Mythos, deemed too risky for general release, have now attracted the very regulatory scrutiny that poses a substantial risk to its business operations.
In an interesting side note, Sam Altman of OpenAI has previously characterized Anthropic’s cautious approach as “fear-based marketing.” He suggested that the promotion of Mythos as an extraordinarily dangerous model could backfire, leading regulatory bodies, including the U.S. government, to take notice. While he didn’t foresee a government-imposed shutdown, Altman’s remarks highlight the potential repercussions of branding AI technology as uniquely perilous—consequences that Anthropic is currently grappling with.
