Anthropic blocks Fable 5 and Mythos 5: impact on AI and european companies

On June 12, 2026, the United States government ordered Anthropic to suspend access to Fable 5 and Mythos 5—its most advanced AI models—for all foreign nationals, both inside and outside the U.S. Since Anthropic could not filter users by nationality, it disabled both models globally within 90 minutes, interrupting service for hundreds of thousands of users, including paying customers. The measure was partially eased on June 22, but the incident has already sparked a debate on AI technology control and its consequences for European companies.

Why did Anthropic block Fable 5 and Mythos 5?

On June 9, 2026, Anthropic publicly released Claude Fable 5, its newest and most advanced AI model. Within days, it became one of the most advanced and high-performing public models in Vals AI tests, a company specializing in standardized model performance benchmarking.

Fable 5 is the public version of Mythos 5, an even more advanced model that Anthropic had distributed in limited form to around 50 organizations—including Amazon, Apple, Google, Microsoft, and CrowdStrike—within the Project Glasswing program, designed to leverage its capabilities for defensive cybersecurity purposes.

On June 12, at 5:21 p.m. Eastern Time, Anthropic received a letter from the U.S. Department of Commerce signed by Secretary Howard Lutnick. The content: immediately suspend access to Fable 5 and Mythos 5 for any foreign national, inside or outside the U.S., including Anthropic’s own foreign employees. The official justification was national security.

Anthropic explained that it could not distinguish users by nationality in real time. For this reason, it chose to block both models for everyone in order to comply with the directive. Access to the company’s other models was not affected.

The non‑universal jailbreak contested by Anthropic

The government’s official motivation refers to an alleged jailbreak of Fable 5—a technique that bypasses an AI model’s safety constraints by convincing it to ignore its internal rules.

Anthropic complied with the directive but publicly challenged its foundations. According to the company’s official blog, the technique identified falls into the category of non‑universal jailbreaks: a technique that works only under very narrow conditions, with prompts crafted specifically for a particular version of the model, and stops being effective as soon as the context changes. The same vulnerabilities, Anthropic argues, can be found in other publicly available models without any bypass technique.

Some independent security researchers who reviewed the report confirmed this assessment, yet the U.S. government still deemed the available material sufficient to impose a total shutdown.

The structural problem: no shared standard

The divergence between Anthropic’s assessment and the government’s highlights a structural issue in the sector. As of today, there is no shared, publicly verifiable technical standard for classifying the severity of a jailbreak on an AI model.

The same technical result—a single test on a limited number of programs, which produced no working exploits—was interpreted by the company as a manageable risk and by the government as sufficient grounds for a global shutdown. Without a public criterion, the final decision depends entirely on whoever enforces it at that moment.

Why does Anthropic’s shutdown worry Europe?

The news prompted immediate reactions from European politicians across very different political backgrounds. According to Euronews, French, British, and Dutch politicians interpreted the case as a sign of Europe’s dependence on technologies controlled elsewhere.

The common thread is clear: relying on technologies that others can revoke without notice exposes European countries to operational disruptions beyond their control. The proposed solution is to invest in AI models and infrastructure developed in Europe.

The European regulatory framework is moving in this direction. The AI Act requires AI model providers to supply technical documentation on capabilities, limitations, and risks, with additional requirements for models classified as systemic-risk. It is an attempt to build a verifiable technical criterion for risk assessment—the very criterion that was missing in the American decision-making process regarding Fable 5.

How risky is it to depend on a single AI platform?

AI is no longer just a tool. It has become a critical part of the digital infrastructure, like energy or the Internet. And like any infrastructure you do not control: if someone else decides to turn it off, you cannot do anything.

That is what happened with Fable 5 and Mythos 5: European manufacturing companies, hospitals, and financial institutions suddenly found themselves without access to the models they were using—not due to a technical issue, nor by their own choice, but because of a decision made across the ocean.

As Marco Montemagno wrote in L’Espresso: “If your assistant, your workflow, your company or your product hangs from the switch of a single platform, you are not truly autonomous. You are a guest.”

Technological sovereignty, in this sense, is not an abstract concept. It means being able to rely on tools that are dependable, available, and compatible with the context in which they operate. Building AI systems that are more portable, independent, and rooted in the European regulatory environment is not an ideological choice: it is an operational continuity choice.

Sources:

• Wired, “Fable 5 and Mythos 5: Why Anthropic Decided to Launch Two New Versions of Its Most Debated Model” (June 13, 2026), https://www.wired.it/article/fable-5-mythos-5-anthropic-nuove-versioni-modello-ai-access-o/

• Euronews, Nathan Rennolds, “Europe’s Reactions to Anthropic’s Shutdown of Fable 5 and Mythos 5” (June 13, 2026), https://it.euronews.com/2026/06/13/sveglia-per-leuropa-reazioni-allo-stop-di-anthropic-a-fable-5-e-mythos-5

• Geopop, Giuseppe Servidio, “Anthropic Disables Claude Fable 5 and Mythos Due to U.S. Government: What’s Happening” (June 15, 2026), https://www.geopop.it/anthropic-disattiva-i-modelli-ai-claude-fable-5-e-mythos-a-causa-del-governo-usa-cosa-sta-succedendo/

• Euronews, Nathan Rennolds, “Why Anthropic Blocks Access to Fable 5 and Mythos 5 AI Models” (June 13, 2026), https://it.euronews.com/2026/06/13/perche-anthropic-blocca-laccesso-ai-modelli-ia-fable-5-e-mythos-5

• Cybersecurity360, Luisa Franchina and Maria Beatrice Versaci, “AI, Export Control and Transparency: Why the Fable 5 Case Also Matters for European Companies” (June 19, 2026), https://www.cybersecurity360.it/nuove-minacce/claude-fable-5-sospeso-per-un-jailbreak-non-universale-perche-conta-anche-per-le-aziende-europee/

• L’Espresso, Marco Montemagno, “What the Dirty Story of the Fable 5 Shutdown Teaches Us” (June 20, 2026), https://lespresso.it/c/economia/2026/6/20/anthropic-caso-blocco-fable-5-amazon/62835

Wondering how to reduce risks linked to dependence on external AI models and build safer, more compliant AI use within your company? Contact us for a consultation.