The White House shelved the security alarm over Anthropic's latest Claude models in less than a month. Commerce Secretary Howard Lutnick informed the company that no license will be required to export or transfer the Claude Fable and Claude Mythos AI systems within the country. The decision abruptly removes the barriers erected after the models were flagged as potential national-security risks.

What changes for Fable 5 and Mythos 5

Fable 5, positioned as an accessible evolution of the Claude family, becomes available today in all markets where Anthropic operates. Mythos 5, designed for more sensitive enterprise and government workloads, follows a different path: access for US organizations was already restored on June 26, but Anthropic is now working with the government to expand it to a "broader set of domestic and international partners" through the Glasswing program. This channel allows cybersecurity researchers at trusted companies to use Mythos for defensive purposes, an area where the model shows specific threat-analysis capabilities.

Export controls and AI: a tension that isn't fading

The rapid U-turn does not erase the political signal. The Trump administration had spotlighted the danger of the most advanced models, igniting a debate that remains open. Export restrictions on sensitive technology are a classic tool of American industrial policy, but applying them to software – and specifically to Large Language Models – introduces unprecedented complexity. Unlike hardware, where physical delivery of chips or servers can be tracked, an AI model can be distributed via API, downloaded as binary weights, or integrated into cloud services without clearing customs.

For those managing on-premise infrastructure, the issue is twofold. On one hand, regulatory uncertainty makes it harder to plan the adoption of external LLMs: the risk of losing access to a model on which time and fine-tuning resources have been invested is not theoretical. On the other hand, the ability to run models locally – when weights are released – becomes again a sovereignty factor, because it allows keeping data and inference within one's own physical and legal boundaries, reducing dependence on cloud services subject to unilateral government decisions.

The trust issue and the role of Glasswing

Lutnick's letter acknowledges that Anthropic has "taken steps in close coordination with the US government to address the risks" posed by the models. This marks a model of public-private collaboration likely to be repeated: companies strengthen safety testing and alignment mechanisms in exchange for an export green light. The Glasswing program embodies this logic: instead of closing access to Mythos altogether, it is reserved for a selected perimeter of users with defensive purposes.

For European organizations working with regulated data (GDPR, residency requirements) this scheme has concrete implications. If access to an advanced model depends on joining a US government-led club, bringing inference on-premise – on owned or rented hardware – becomes a tool for autonomy but also for compliance: it avoids routing prompts and data through endpoints controlled by non-EU jurisdictions. From this perspective, the global release of Fable 5 adds a piece to hybrid architectures where self-hosting is paired with as-a-service models, while Mythos remains a tightly regulated asset.

Beyond the Anthropic case

The episode shows how porous and subject to sudden revisions the boundary between national security and AI access is. Those designing infrastructure for on-premise Large Language Models – whether banks, public administration, or critical manufacturing – can no longer ignore the geopolitical variable when choosing models. The fact that restrictions are lifted today offers no guarantees for the future, and actually pushes toward architectures that make the inference engine interchangeable without rewriting pipelines and integrations. The path taken by Anthropic with the US government suggests that the most capable models will remain at the center of a constant negotiation between innovation and control.