On June 12, 2026, the US Department of Commerce issued an export-control order directed at Anthropic, requiring it to immediately suspend access to its two top-tier models, Claude Fable 5 and Claude Mythos 5, for any foreign national inside or outside the country, including Anthropic's own employees who are not US citizens. The measure, reported by the press as having been communicated to CEO Dario Amodei and invoking national-security authorities, led the company to block all public access to both models on a global scale. The rest of the Claude family —Opus, Sonnet and Haiku— continues to operate normally.
What the government says, and what Anthropic answers
According to the official account reported by the media, the trigger was the alleged detection of a method to bypass Fable 5's safeguards through jailbreaking, which prompted the invocation of national-security powers. The order, however, arrived without a public technical description of the specific risk. Anthropic has responded with a direct criticism of the procedure: it argues the government should be able to block unsafe deployments, but "through a transparent, fair, clear and technically grounded legal process", and that this action does not meet those principles. The company has also noted that rival models with equivalent capabilities have not faced comparable restrictions, a point that fuels the debate over the proportionality of the measure.
These claims are best treated as what they are: the positions of the parties. The government's appeals to a security risk it has not disclosed in detail; the company's, to a flaw in process and consistency. Without the underlying technical documentation, neither is independently verifiable as of today.
The mechanism: export controls applied to software, not hardware
What is technically novel is not the block itself, but the tool. US export controls have for years been applied to hardware —chips, manufacturing equipment— but applying them to access to an AI model served via API extends that logic to software as a service. In practice it means a trained, hosted model can become unavailable not through a technical fault or a commercial decision, but through an administrative directive; and the "any foreign national" scope forces the company into a global shutdown, because there is no clean way to segment access by nationality without breaking the service for everyone.
The European reaction: a wake-up call
In Europe, the news has been read as a warning. Several politicians have called it a "wake-up call" and demanded greater investment in homegrown technology, with a message that captures the heart of the matter: a nation that depends on others for its technology is a nation that can be "unplugged" overnight. The episode also lands amid Europe's push for technological autonomy —from the EuroHPC AI Gigafactories to the technological-sovereignty package and the Chips Act 2.0— and gives ammunition to those who argue that sovereignty is measured not only in compute capacity, but in real control over the full stack.
The engineering lesson: diversify or stay exposed
Beyond the geopolitical noise, there is an operational conclusion that affects any technical team. Any organisation that has built agentic workflows or production applications tied exclusively to a single closed-API provider is exposed to immediate disruption if that provider faces an injunction, a cyberattack or, as here, an export-control directive. The recommendation analysts keep repeating is clear: diversify AI providers, abstract the model layer, and avoid critical dependence on a single model or jurisdiction. It is not a question of brand, but of business continuity.
A dependence that is no longer theoretical
For years, the risk of "having your model switched off" was treated as a whiteboard scenario. The Fable 5 and Mythos 5 case turns it into precedent: a regulatory change in one jurisdiction can simultaneously cut off users and companies around the world. For Europe, the technical message is consistent with the one we keep making: resilience is not bought with compute power alone, but by designing systems that do not break when a single piece —for commercial or political reasons— stops being available.
It remains to be seen how the procedure evolves and whether the restriction holds, narrows or is reversed. But the architecture lesson is already on the table, and it does not depend on the outcome.
Back to publications