Export Controls Lifted

On June 12, the US Department of Commerce ordered Anthropic to pull its two most advanced AI models, Claude Fable 5 and Mythos 5, from global access over national security concerns. On June 30, less than three weeks later, the department lifted the restrictions entirely. The models will be available again starting Wednesday.[1]

The episode is a useful case study in how AI governance actually works in practice, as opposed to how it is discussed in policy papers.

Fable 5 is Anthropic's consumer-facing reasoning model, capable of complex multi-step tasks. Mythos 5 is the enterprise variant, designed for cybersecurity work, including identifying and exploiting vulnerabilities in code. Both were released on June 9. Three days later, the government forced a global suspension after discovering what it characterized as a method to bypass the models' safety restrictions, a jailbreak.[2]

Anthropic's response at the time was notably blunt. The company said it disagreed that "the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people."[2]

Commerce Secretary Howard Lutnick's June 30 letter to Anthropic said the company had "agreed to proactively detect and address security risks associated with the models" and would collaborate on future releases. The department reserved the right to reimpose restrictions if necessary.[1]

Rival OpenAI has also been subject to similar pressure. Its GPT-5.6 model was released only to a limited set of approved partners. OpenAI's CEO Sam Altman posted on X that "this isn't quite the process that we think is optimal," a diplomatically worded complaint about a process that took his company's flagship model off the open market.[3]

What is striking here is the speed. Three weeks from global ban to full restoration. The government did not hold public hearings. There was no formal rulemaking. A letter from the Commerce Secretary was the entire mechanism. For an industry that regularly discusses regulation in terms of years-long legislative processes, the actual intervention was remarkably fast and remarkably informal.

Anthropic has called for a standardized framework to assess critical vulnerabilities in advanced models, working with Amazon, Microsoft, and Google on the effort. This is a sensible response, and it tells you something about where the company thinks the real risk lies: not in the models themselves, but in the ad hoc process that can shutter a product for weeks with no clear criteria for restoration.[1]

The broader question is whether this approach scales. Frontier models are getting more capable. The stakes of a jailbreak in a model that can find zero-day vulnerabilities are not trivial. But a process where the government imposes a ban, the company negotiates in private, and three weeks later everything is restored does not inspire confidence that anyone is doing rigorous safety analysis. It looks more like a fire drill than a framework.

The Trump administration's June 2 executive order called for voluntary frameworks for testing and releasing frontier models over a two-month window. That deadline is approaching. Whether the result is something meaningful or just a checkbox remains to be seen. The Anthropic episode suggests the government is learning by doing, and the companies are learning that compliance, even hurried compliance, gets you back online fast.[3]

For everyone else, the takeaway is simpler: your access to a frontier AI model can disappear and reappear in under a month, based on a process that has no public documentation. Plan accordingly.