Anthropic disables two models under US order

Anthropic said on June 12 that it had received a US government export control directive requiring it to suspend access to Claude Fable 5 and Claude Mythos 5 for every foreign national, including people inside the United States and foreign national Anthropic employees. Because the company says it cannot apply that filtering immediately without risking non-compliance, it has disabled both models for all customers. Other Claude models are not affected, according to the same statement.

The point matters because it shifts the frontier AI policy debate. Until now, the most visible restrictions have focused on chips, chipmaking equipment, and some cloud services. In this case, the controlled object is a commercial AI model that was already available or close to available. Fable 5 had been introduced a few days earlier as a Mythos-class model intended for broad use, while Mythos 5 was limited to Project Glasswing partners working on software vulnerability discovery. A vulnerability, in this context, is a weakness in software that could be exploited to compromise it.

Anthropic says the government letter did not provide specific details about the national security concern. The company says its understanding is that the government believes there is a method for bypassing, or jailbreaking, Fable 5. A jailbreak is an attempt to make a model produce an answer that its safeguards would normally block. Anthropic disputes the scale of the issue: it says it reviewed a demonstration tied to a small number of previously known, minor vulnerabilities, and it points to thousands of hours of red-team testing with internal teams, government bodies, and private organizations before launch.

For developers, the practical lesson is less flashy than a new benchmark but more important for planning. A frontier model provider can now see a model abruptly disabled for sovereignty and security reasons, even when customers have not changed their behavior and even when the product was meant to serve a global market. Companies building on a single frontier model need to account for political availability risk, alongside technical outages, pricing changes, and contract limits. For governments, the episode also exposes a difficult boundary. Controlling models directly may reduce some risks, but it can also interrupt defensive research, software security work, and legitimate use by foreign partners. What happens next will depend on how quickly Anthropic and US officials define the conditions for restoring access.