In recent years, the strategic contest over advanced artificial intelligence has been defined above all by the struggle to control the hardware foundations of frontier AI.
The key policy question was how states could control the physical inputs that enabled frontier AI, such as advanced chips, semiconductor manufacturing equipment, and the supply chains that support them.
Export controls were therefore built around a relatively clear assumption. If a rival could be denied the hardware required to train the most capable models, its progress at the frontier could be slowed.
US export controls on China were largely designed around this logic, seeking to limit China’s access to the advanced hardware and production capabilities needed to compete at the AI frontier.
The export control directive affecting Anthropic’s Fable 5 model suggests that this logic has entered a new phase.
The object of control now extends to the model itself, to its operational capacities, and to the channels through which users may interact with those capacities.
In other words, the frontier has moved from the factory floor and the data centre into the access layer of the model economy.
Related
This shift makes it increasingly difficult to think of the advanced AI market as an ordinary global technology market, where any sufficiently resourced customer can buy access to the best available systems.
At the highest tier of capability, access is becoming subject to security classification, alliance politics, and direct state intervention.
The implications are especially significant for third countries. For states that claim strategic autonomy, AI can no longer be treated as a service to be purchased abroad when needed.
Domestic compute capacity, local model development, and a deep research base are becoming core components of national power.
Anthropic’s rise
Anthropic’s recent trajectory captures this transformation with unusual clarity. The company has become one of the most closely watched firms in the AI sector, but its rise has also acquired a sharper national security dimension.
Its Claude models have been used on classified Pentagon networks since 2025, yet the company’s reservations about certain military applications, including domestic surveillance and fully autonomous weapons, created tension with the Department of Defense.
That tension hardened into a more serious institutional conflict in February 2026, when the Pentagon classified Anthropic as a supply chain risk.
Anthropic now stands at the centre of another security controversy, this time involving the Claude Mythos 5 and Fable 5 models.
Mythos 5 appears to have been treated as exceptional because of its performance in domains such as cybersecurity, biology, and health.
Rather than releasing it for general use, Anthropic placed it in restricted-access programmes such as Project Glasswing, limiting availability to a trusted set of users.
Fable 5 was designed as a more commercially usable version of the same underlying architecture, with additional safeguards intended to make its capabilities available in a more controlled product form.
Anthropic sought to convert Mythos-level reasoning, coding, visual comprehension, and long-context capacity into a safer and more scalable service.
Sensitive requests in cyber and biological domains would be routed away from Fable 5 toward the more limited Opus 4.8 model to reduce exposure in areas considered especially dangerous.
That opening lasted only a few days. On June 12, Anthropic announced that it had received an export control directive from the US government requiring it to suspend access to Fable 5 and Mythos 5 for non-US persons.
The company said the government had not provided a detailed written justification, but had presented oral evidence suggesting that Fable 5 could be bypassed through a relatively simple jailbreak technique. Anthropic complied and cut off access.
The restriction covered foreign nationals based in the US and even foreign employees of Anthropic itself.
Because the company could not quickly and reliably determine who should and should not retain access under these rules, it disabled the models for all customers.
The new direction of export control
US AI policy in recent years has focused heavily on advanced chips, semiconductor manufacturing equipment, hyperscale data centres, and high-capacity compute infrastructure.
That approach rested on the premise that restricting the means of AI production would constrain the ability of rival actors to train the most advanced models. Chips were treated as the industrial foundation of AI power.
If access to that foundation could be limited, model development elsewhere would slow.
Until recently, the most sensitive non-hardware element in this framework was the model weight. Model weights can be understood as the trained parameters that carry a system’s capability.
Once the weights of an advanced model are copied, the recipient can run the model without depending on the provider’s infrastructure.
It may remove safety layers, retrain or fine-tune the system for its own purposes, transfer capabilities into smaller models, or integrate the system into military, intelligence, cyber, or surveillance platforms.
For that reason, model weights are better understood as portable strategic capacity than as ordinary software files. Conversely, in the API model, the user does not possess the weights.
