Anthropic disables Fable 5 and Mythos 5 after U.S. export directive

Anthropic has cut off access to its Fable 5 and Mythos 5 systems after the U.S. government ordered the company to suspend use by any foreign national, including Anthropic employees overseas and in the United States. The company said it received the directive at 5:21 p.m. ET on June 12 and, to comply, disabled both models for all customers while leaving its other products untouched.

The move puts a sharp question at the center of AI oversight: when does a disclosed safety weakness become a national-security event? Anthropic said the government did not spell out the threat, but its understanding was that officials believed they had found a way to bypass, or “jailbreak,” Fable 5. The company described the issue as a narrow, potential jailbreak and said that did not justify pulling a commercial model used by hundreds of millions of people.

Anthropic also argued that the alleged capability was not unique. It said the same kind of behavior was already available in other public models, including OpenAI’s GPT-5.5, and that such techniques are routinely used in defensive cybersecurity work. That framing matters because the government directive appears to treat access by foreign nationals as an export-control problem, placing model access in the same policy bucket as technology transfer and other sensitive national-security concerns.

The clash lands awkwardly for a company that has built part of its brand around publishing its own weaknesses. Anthropic has repeatedly promoted its Responsible Scaling Policy and said its most powerful systems carry ASL-3 protections. It has also worked with the U.S. AI Security Institute and the U.K. AI Security Institute on red-teaming and safeguard testing, including research on universal jailbreaks that, the company said, led it to fundamentally restructure its safeguard architecture. Its earlier work on many-shot jailbreaking and constitutional classifiers was meant to harden defenses, not hide the underlying risks.

That history now raises a harder question for the industry: whether transparency about model vulnerabilities will be read as responsible disclosure or as evidence that a system has crossed a regulatory threshold. Anthropic said it believes there may be a misunderstanding and said it is working to restore access, but the case suggests future safety reporting may be shaped as much by export law and enforcement pressure as by technical merit.