United States: Anthropic Forced to Limit Fable 5 and Mythos

Washington Imposes Drastic Restrictions on Fable 5 and Mythos 5

Just days after the launch of Fable 5, an artificial intelligence model from the Mythos series aimed at the general public, the United States has decided to restrict access to it. This decision, motivated by national security concerns, affects all foreign users. Anthropic, the company behind these models, responded by calling the situation a "misunderstanding" and is committed to restoring access quickly.

A Security Directive Impacting All Users

Claude users were surprised to discover that Fable 5 was "currently unavailable." This suspension follows an official directive received by Anthropic on June 12, 2026, citing potential threats to national security. The U.S. government demands the suspension of access to Fable 5 and Mythos 5 for all foreigners, whether they are in the United States or elsewhere, including foreign employees of Anthropic.

The directive does not specify the exact nature of the threat. However, Anthropic mentioned a suspicion of jailbreak, a technique that allows bypassing the model's protections. Unable to guarantee the exclusion of foreign users, Anthropic chose to disable access for all its users.

For context, Claude Mythos was introduced in April 2026 but was only accessible to about fifty partners as part of the Project Glasswing, due to its cyber capabilities deemed sensitive. On June 9, Anthropic launched Fable 5 for the general public, with safeguards redirecting sensitive requests to Claude Opus 4.8. Mythos 5, without these restrictions, remained reserved for Glasswing partners.

Anthropic Complies While Contesting the Directive

Although Anthropic has complied with the directive, the company contests its legitimacy. In a statement, it raised several points:

• Fable 5 underwent thousands of hours of security testing, involving the U.S. government, UK AISI, and other organizations, with no universal jailbreak detected.
• The identified vulnerability is limited and not universal, consisting of asking the model to correct flaws in a codebase, a common practice in cybersecurity.
• Similar capabilities already exist in other public models, such as GPT-5.5 from OpenAI, and are not exclusive to Fable 5.
• No concrete examples of harm caused by this technique have been provided to Anthropic.
• A 30-day data retention policy was implemented at launch to detect and correct any circumvention.

Anthropic states:

"We comply with the government directive by removing access to Fable 5 and Mythos 5 for all users. However, we contest that the discovery of a potential security flaw, even a minor one, justifies the withdrawal of a commercially deployed model at scale. If this standard were applied across the industry, it would effectively block any new model deployment for all innovative providers."

The company is actively working to restore access and promises to provide new information within 24 hours.

Europe Faces Its Technological Dependence

This American decision reignites the debate on the dependence of states, particularly European ones, on American technologies. It demonstrates that a government can, overnight, cut access to essential digital tools for foreign users. On X, the French political spectrum expresses concern about the implications for digital sovereignty.

Benjamin Haddad, the French Minister Delegate for Europe, stated that this decision by the Trump administration accelerates the geopolitical battle over AI. He emphasized that Europe cannot simply be an open market dependent on technologies designed and controlled elsewhere. Anne Le Hénaff, the Minister for Digital Affairs, added that major powers now view strategic technologies as instruments of power, calling for increased European investment in AI.

Benjamin Haddad warns:

"Let's not kid ourselves: whether Democrats or Republicans, Donald Trump's successors will not reverse this type of decision. Control over models is an acceleration of Biden's practice of export restrictions on chips (electronic chips essential for the functioning of AI models, editor's note)."