On June 13, 2026, the Trump administration issued an export control directive that forced Anthropic to disable two of its most advanced models for all customers worldwide.
The models involved: Fable 5 and Mythos — the most powerful in Anthropic's lineup.
The cause: Amazon.
How It Happened
On the evening of Thursday, June 12, Amazon contacted Trump administration officials directly, sharing an internal report showing that Amazon's security researchers had managed to jailbreak the Mythos model to access capabilities the company considered a national security risk.
Amazon CEO Andy Jassy was personally in touch with administration officials on the matter.
The Department of Commerce responded Friday with a formal directive: no foreign citizen — including Anthropic's own employees — would be allowed to access Fable 5 and Mythos.
Anthropic had to disable access for all customers, without exception, in order to comply.
Anthropic's Response
Anthropic said it had to "abruptly disable" the models to comply with the order. But not without stating its own position:
The company believes a partial, limited-scope jailbreak doesn't justify pulling a commercial product. In its view, the government's response is disproportionate to the specific nature of the vulnerability found.
It's a delicate position: Anthropic can't ignore a directive from the US federal government, but it also doesn't want the market to perceive its models as inherently dangerous.
The Context: AI Safety as a Political Battleground
This episode fits into a broader context. The Trump administration has shown growing interest in controlling who accesses the most advanced AI models developed in the US — arguing that frontier model capabilities could be used by foreign actors (China in particular) for cyberattacks, disinformation, or military applications.
The logic mirrors export controls on advanced chips (the restrictions on NVIDIA's sales to China), now applied to the AI models themselves.
The difference from chips is that an AI model distributed via API is far harder to "control" than a physical product. The jailbreak Amazon demonstrated shows that the line between a commercial application and a potentially sensitive one isn't always clear-cut.
What This Means for Anthropic Users
If you're using Fable 5 or Mythos in your applications or workflows — whether you're a developer, a business, or a researcher — access isn't available until the situation is resolved.
For Anthropic's earlier models (like Claude Sonnet and previous versions of Opus), there are currently no similar restrictions.
The situation is evolving fast: dialogue between Anthropic and the administration remains open, and there could be developments in the coming weeks.
What's certain is that the AI industry just got a concrete demonstration of how government policy can cut off access to commercial products overnight. It's a risk that anyone building critical applications on frontier models needs to factor into their architecture — both in terms of fallbacks and provider diversification.