Washington Lifts Export Controls on Anthropic's Fable 5 and Mythos 5

For two weeks, no user outside the United States could legally access Anthropic's most powerful models. On June 12, the US government ordered the company to immediately suspend access to Claude Fable 5 and Claude Mythos 5 for all foreign nationals, citing national security authority. The controls were lifted on June 30, and on July 1 Anthropic restored Fable 5 globally, with a new safety filter and tighter cooperation with Washington.

What was discovered

The trigger was a report from Amazon researchers who found a way to bypass Fable 5's safeguards using a carefully constructed prompt. The model could be coaxed into identifying specific software vulnerabilities and describing methods to exploit them, exactly what its built-in safety barriers were meant to prevent.

The government deemed this sufficient grounds to block foreign nationals from accessing the model, both those abroad and those on US soil. The decision covered Fable 5, the publicly available version with safeguards, and Mythos 5, the same architecture without cyber restrictions, made available only to cybersecurity specialists under the Project Glasswing program.

Two weeks of negotiations

US Commerce Secretary Howard Lutnick wrote on X that the government spent two weeks working with Anthropic to align Fable with US security interests. The result is a new safety classifier that blocks the specific technique described in Amazon's report in more than 99 percent of cases, rerouting blocked queries to the older Claude Opus 4.8 model.

Anthropic also widened its so-called safety margin, meaning the filter now blocks more harmless queries than before as well. The company confirmed at the same time that testing had revealed an identical vulnerability in other models, including OpenAI's GPT-5.5 and Kimi K2.7, suggesting the problem was not unique to Fable 5 but affected an entire class of today's systems.

Shared industry standards

In response to the incident, Anthropic pledged deeper cooperation with the government and, together with Amazon, Microsoft and Google, said it would develop a shared framework for assessing the severity of jailbreaks. The four criteria in that framework are the capability gain a model exhibits once its safeguards are bypassed, the breadth of that gain, how easily the vulnerability can be weaponized, and how easily third parties can discover it.

It is the clearest signal yet that the industry's major players want a common language for describing serious model-safety incidents, rather than assessing each case in isolation. For Washington, it is also a way to selectively open global markets to labs that can demonstrate real safeguards, while maintaining a hard line against China.

What it means for the industry

The episode comes as open Chinese models such as GLM-5.2 are nipping at the heels of Western leaders at a fraction of the price. Analysts cited by financial media note that the swift reinstatement of Fable 5 for foreign markets gives Anthropic an edge in the race for institutional and research customers who had held off on purchasing decisions pending clarity on the legal situation.

For companies and institutions in Poland using Claude via the API, this marks the end of two weeks of uncertainty, but also signals that access to the most powerful models can be temporarily suspended at any time by administrative decision in the US, regardless of where the user is located. Companies planning deployments built on Anthropic's latest models should factor in this regulatory risk when designing critical system architecture.

Sources: Redeploying Claude Fable 5 (anthropic.com), Anthropic restoring access to its most powerful AI models signals a necessary truce with the U.S. government (fortune.com), Trump Just Cleared Anthropic to Go Global, and Google DeepMind Should Be Sweating (247wallst.com)