After the United States Government lifted export restrictions at the end of June, Anthropic has restored access to two advanced models. Under corporate arrangements, Claude Fable 5 was re-opened to the world on 1 July, while Claude Mythos 5 was re-opened only to approved United States institutions.

Security incidents triggered restrictions in June

This interruption began on June 12. At that time, the United States imposed export controls on both models. As the relevant order came into effect immediately and Anthony was unable to verify in real time the nationality of all users, the company eventually suspended all user visits to avoid violations.

This restriction is related to a cyber-security discovery. Amazon researchers have previously reported that Fable 5's security protection can be bypassed, which is commonly referred to as a “breakout”. In the testing, the model is induced to identify software gaps and, in one case, examples of codes that can be used to exploit them.

There are significant differences in restoration arrangements

Anthropic states that the United States Government lifted export restrictions on Fable 5 and Mythos 5 on 30 June. Of these, Mythos 5 was based on the same bottom model as Fable 5, but with fewer security restrictions, the restoration was narrower and limited to only part of the United States.

  • Fable 5: July 1st Global Visit resumes
  • Mythos 5: Recovery to approved U.S. agencies only
  • Export restrictions: formally lifted on 30 June

The company will expand its cooperation with the government.

While restoring services, Anthropic is also strengthening its collaboration with the United States Government. The company stated that prior to the publication of the model, advance access to the front-line model and its security protection would be provided to designated government agencies and threat intelligence would be shared.

Anthropic is also working with Amazon, Microsoft and Google on a framework to assess the extent of the risk of “breakout”. The arrangement echoes the AI Security Executive Order issued by the United States on 2 June and also shows that government reviews are becoming normal before the front-line model goes online.