Tomorrow, will you have to show your passport to use AI?

TL;DR

Artificial intelligence is generally analyzed from the angle of the power of the models, the colossal investments they require or the competition between laboratories. However, another mutation is perhaps taking place, discreet but potentially more structuring, namely the control of access to the models themselves.

The agreement which has just been concluded between Anthropic and the Trump administration constitutes, in this regard, a precedent. After two weeks of tensions, during which Europe but also Silicon Valley discovered a new sovereignty problem, the US Department of Commerce authorized the gradual restoration of access to the Mythos 5 model for some of the company’s customers. In return, Anthropic has committed to strengthening the safeguards of its models and working with Washington on protocols governing future generations of artificial intelligence.

If the episode can be read as a simple compromise, it actually reveals a much more significant evolution: the governance of border models is no longer just a matter of laboratories, and is gradually becoming a subject of industrial policy, national security and de facto technological sovereignty.

The Anthropic affair marks the end of an era

This new agreement would therefore authorize Mythos 5 to be gradually put back into service with a portion of authorized customers, while negotiations continue regarding Fable 5.

In a letter to Anthropic, Commerce Secretary Howard Lutnick reportedly indicated that the company is committed to working with the US government on protocols for future models.

Until now, the release of a new model was almost exclusively the responsibility of internal laboratory teams: research, security, red teaming, alignment, then marketing; from now on, the American administration intends to participate upstream in this process.

This precedent goes far beyond Anthropic, OpenAI announced that GPT-5.6 Sol would, initially, be reserved for a limited number of organizations approved by the American administration, Meta is invited to voluntarily submit its models for federal evaluations, Google, Microsoft, xAI and Anthropic are already participating in the work of the Center for AI Standards and Innovation, responsible for evaluating the most advanced models.

For several months, laboratory leaders, cloud managers, cybersecurity specialists and representatives of the American administration have been discussing, often informally, the need to more strictly regulate access to border models. These discussions, until now largely invisible, are now notorious and the Anthropic affair offers the first concrete example, with the central question of who can have access to their models, particularly the most advanced.

Towards “KYC” for artificial intelligence

The idea is not revolutionary and already exists in other industries. Banks apply Know Your Customer (KYC) procedures to identify their customers, cloud providers rely on sophisticated identity and access management systems, defense sectors have long operated with clearance levels.

Artificial intelligence could also adopt these mechanisms. Thus, access to the most powerful models would no longer be universal, and depending on the level of risk associated with a model, the user may have to justify their identity, their organization, their country of origin, their activity and sometimes even the purpose of their project.

A model specialized in offensive cybersecurity, computational biology or autonomous agents could require more or less strict authorization depending on the regulations.

The principle would not consist of prohibiting models, but of creating several levels of access, comparable to the levels of authorization already used in administrations or critical infrastructures.

Practical aspect but not the least, this development does not suppose any technological breakthrough, the infrastructures already exist. For example, Microsoft which leverages Azure Government, Azure Secret and Microsoft Entra to manage identities and access in sensitive environments, Amazon has GovCloud, Secret Region and all of its IAM services, Google offers Cloud Identity and Assured Workloads for regulated environments.

In other words, the necessary technical building blocks are already operational, and all that remains is to apply them directly to artificial intelligence models.

The distribution cycle of a model would gradually evolve from a “release first” model towards a “clearance first” logic, a change to which AI laboratories and players would have to adapt.

A new layer of control is emerging

This development would profoundly modify the balance of power in the industry, because if the laboratories will remain masters of their models, the hyperscalers could become the real guardians of their distribution.

Microsoft, Amazon and Google, which already master user identification, authorization management, audit logs, sovereign environments and certified infrastructures for administrations or defense, could be called upon to carry out this control.

This logic gradually brings the frontier models closer to the regime of dual technologies. They cease to be simple software to become critical infrastructures, in the same way as certain space, nuclear or military capabilities.

Large groups will probably have to integrate new governance of access to models, where information systems departments will have to administer specific authorizations, log uses, audit interactions with certain models and demonstrate their compliance during controls.

A new market could thus emerge around “AI Identity Management” and “AI Access Management”, at the crossroads of cybersecurity, regulatory compliance and the cloud.

This development also illustrates American pragmatism in terms of AI governance. Rather than concentrating regulatory effort on the transparency of models or on the documentary obligations imposed on developers, Washington seems to favor a more operational lever: control of the distribution and conditions of access to the most advanced models.

Tomorrow, models will perhaps be designed to satisfy regulators

A consequence is still largely absent from the debates, because if market access tomorrow depends on government validation, laboratories will have to integrate this constraint from the design of their models.

Safety and alignment teams will no longer work only to reduce hallucinations, limit malicious uses or improve the robustness of systems, but will also have to meet regulatory requirements likely to condition marketing authorization.

In other words, future models could be designed not only to maximize their performance, but also to meet compliance criteria defined by the authorities. So the frontier models will perhaps no longer be just “safety by design”, but will have to be “regulation by design”.

A change which could prove even more structuring than the next technological advances. For two years, global competition has been analyzed as a race for parameters, GPUs and gigawatts; the Anthropic affair suggests that another battle is beginning, that of controlling access to models, particularly the most advanced.