Washington bans Mythos 5: America wants to control AI models, how will the industry respond?

Late on Friday, the US government ordered Anthropic to suspend access to Mythos 5 and Fable 5, its most advanced cybersecurity models, with immediate effect. The directive is based on national-security considerations and Washington’s export-control powers.

Anthropic announced the global deactivation of both models a few hours later. The company said the decision followed the identification of a possible mechanism for bypassing Fable 5’s safeguards. The lab nevertheless disputes the seriousness of the evidence presented by US authorities, arguing that the capabilities in question are already available through other publicly accessible models.

The intervention is unprecedented. Since the beginning of the generative AI cycle, Washington has primarily sought to control infrastructure: semiconductors, manufacturing equipment and computing capacity. With Mythos 5, the restriction applies directly to a model. The decision could mark the beginning of a new phase in AI governance and force laboratories to rethink their relationships with governments.

Mythos 5, the first recall of a frontier model

The abrupt nature of the intervention surprised those who have followed debates over frontier-model safety for years. The precedent is considerable. The industry has already faced parliamentary inquiries, export restrictions on sensitive technologies and debates over the responsibilities of model providers. But no government had ever demanded the immediate withdrawal of an already deployed frontier model.

The implications extend far beyond Anthropic. What matters now is not simply the model concerned, but the principle that has been established. For the first time, a state is asserting its authority to intervene directly in the distribution of a capability produced by an artificial intelligence system.

Has Washington imposed an operating licence on AI laboratories?

One of the most underestimated consequences of the Mythos affair could be the accelerating convergence between major US laboratories and the federal government.

For several years, Anthropic, OpenAI and Google DeepMind have conducted growing numbers of red-teaming exercises with government agencies, formed partnerships with national-security bodies and developed oversight mechanisms for their most advanced models. Until now, these arrangements were primarily concerned with risk management. The Mythos episode could give them another function: securing a form of implicit licence to operate.

A new reality is taking shape behind the suspension of Anthropic’s model. The challenge is no longer simply to build the most capable systems, but to demonstrate that they can be deployed without conflicting with US strategic interests.

This development is gradually bringing AI laboratories closer to industries already familiar with this type of relationship with the state: defence, nuclear energy and aerospace. For years, Silicon Valley cultivated the idea that it could move faster than institutions. Part of its future may now depend on its ability to move alongside them.

Europe discovers a deeper dependency affecting its development capabilities

The signal is equally powerful for Europe. The European debate over digital sovereignty has focused on infrastructure, data localisation, secure cloud services and related issues. Mythos reveals a different form of dependency. Ultimately, it does not matter where the user is located, where the company operates or even where its data is stored. When a critical model is developed, hosted and governed under US jurisdiction, access to it remains subject solely to American decisions.

This is probably the most important lesson from the episode for European policymakers. Europe is discovering that its dependency no longer concerns digital infrastructure alone, but cognitive capabilities. As these capabilities become strategic components of economic competitiveness, control over them takes on an importance comparable to that historically associated with energy, telecommunications and semiconductors.

The case for European models is no longer based solely on political sovereignty. It is gradually becoming a matter of operational continuity.

By seeking to control closed models, Washington could strengthen open source

The paradox is that this affair could produce precisely the opposite effect from the one intended. The more governments seek to regulate proprietary models, the more attractive open models become.

If a model hosted by a centralised provider can be suspended, an open-weight model becomes far more difficult to remove from circulation once it has been distributed. This reality was already central to the debate between closed and open-source models. The Mythos restriction gives it a geopolitical dimension.

For advocates of openness, the episode provides further evidence that critical cognitive infrastructure cannot depend exclusively on a handful of companies subject to the decisions of a single state. For governments, by contrast, it reinforces fears of the uncontrolled proliferation of sensitive capabilities.

Behind the technical debate lies a broader question: will the future of AI be organised around licensed and supervised models, or around widely distributed capabilities that are difficult to control?

The real issue: AI’s future markets

It would be a mistake to treat Mythos as an isolated cybersecurity case.

Cybersecurity is probably only the first field in which model capabilities have reached a level of strategic value sufficient to attract government attention. Tomorrow, the same debates will emerge around drug discovery, materials research, energy-system design, aerospace, defence and intelligence.

The issue is therefore not merely technological, but economic. Until now, a model’s value has been measured through its general performance. A different logic is gradually emerging: a model’s value could soon be determined by its ability to accelerate a specific economic or strategic activity.

The industry is therefore entering the era of “capability models”, systems whose primary asset is no longer general intelligence, but the operational advantage they create within a particular field.

Should investors begin pricing sovereign risk?

This development could also change how markets value major AI laboratories.

After debates over revenue, future margins, access to compute, talent and market share, the Mythos affair introduces a new variable: sovereign risk. The more strategically important models become, the more their value will depend on their remaining authorised for use.

This represents a major shift in perspective. Until now, frontier laboratories have been analysed as the software giants of the future. They may gradually come to be viewed as companies operating in regulated industries, where the relationship with the state constitutes an economic asset in its own right.

The real question is no longer whether Mythos 5 will be reactivated, but how the entire ecosystem will respond to this precedent. Once a government demonstrates that it can suspend a frontier model, the industrial strategies, investment flows and technological choices of an entire industry are bound to change. The first responses, or at least the first reactions, should not be long in coming.

Asked about Anthropic’s decision to restrict access to its most advanced models to US users, Anne Le Hénanff, France’s Minister Delegate for Artificial Intelligence and Digital Affairs, said the episode confirmed that artificial intelligence had entered a new phase of strategic competition between technological powers. She also viewed it as further justification for the efforts undertaken by France and the European Union to strengthen their technological autonomy and develop their own infrastructure, computing and innovation capabilities.

“Following the restrictions on chip exports, the United States’ decision to reserve access to Anthropic’s latest AI models exclusively for American users sends a clear message.

The major powers now regard strategic technologies as instruments of power. This reality reinforces Europe’s decision to build its technological autonomy.

France and Europe have considerable strengths to draw upon. Thanks to abundant, low-carbon and competitively priced electricity, we are a natural home for the data centres and infrastructure that will power the AI of tomorrow.

Far from making me doubt our direction, this decision strengthens my determination to accelerate the efforts pursued relentlessly since 2017: more investment, less fragmentation, greater simplification and massive support for our European companies.

And I am proud to be European. Because we believe in free trade and openness, but also in respect for our values, the rule of law and common rules freely established.

More than ever, European digital sovereignty is a necessity.”

Paul Midy offered a similar response:

“This is a clear signal. We cannot depend on the United States for AI and the technologies of the future. We must regain our technological sovereignty. It is possible, but we must commit the necessary resources and energy. Some will lead this revolution, while others will miss it. Our wages, prosperity and ability to defend ourselves will depend on it. We must fight for it.”