When Washington pulled a frontier model offline for foreign nationals, it did more than enforce an export control. It began the territorialization of artificial intelligence, drawing national borders around a technology the internet age assumed would have none.
Rahul PAWA | X- @imrahulpawa
The early internet was sold as a challenge to sovereignty. Information would route around borders, jurisdiction would dissolve into bandwidth, and the territorial state, the durable invention of the seventeenth-century Westphalian order, would find itself outflanked by a global commons it could neither map nor police. For three decades that was the governing assumption of digital policy: code travels, and states adapt. On the evening of June 12, the United States quietly inverted it. It ordered a commercially deployed frontier AI model off the market on national security grounds, and in doing so signalled that the most consequential layer of the digital economy will be organized not as a borderless commons but along national lines.
The mechanics were abrupt. At 5:21 p.m. Eastern Time, Anthropic received an export-control directive instructing it to terminate access to its two most capable models, Fable 5 and Mythos 5, for all foreign nationals, whether located abroad or inside the United States, and including the company’s own non-citizen employees. The directive, signed by Commerce Secretary Howard Lutnick with input from the Bureau of Industry and Security, invoked national security authorities but offered little public technical justification. Because Anthropic could not reliably separate foreign nationals from everyone else in real time, the practical effect was a global shutdown. Within hours both models were offline worldwide. Other Claude models, including Opus 4.8, remained available.
The proximate trigger was a reported jailbreak: a prompting technique that allegedly exposed Fable 5’s restricted cybersecurity functions, the offensive vulnerability-discovery capabilities Anthropic intentionally isolates behind safety controls. The company’s said that the demonstration surfaced only a handful of previously known, low-severity flaws, and that comparable outputs were already obtainable from other publicly accessible frontier models without any special bypass. Anthropic complied while calling the decision a misunderstanding, warning that the same standard, applied consistently, would render future frontier deployments practically impossible.
Whether the bypass was trivial or grave is, for the historian of governance, secondary. What matters is the precedent the state has now set in public. For the first time, Washington has treated a deployed AI model not as a software product but as a controlled strategic asset whose distribution to foreign persons can be halted by executive action.
That is a change in kind, not degree. The closest analogies are not social platforms or cloud services but uranium enrichment technology, jet engines, lithography machines, and advanced semiconductors. Washington already restricts the chips used to train frontier models; it has now moved one layer higher in the stack and claimed authority over the cognitive capabilities those chips produce.
This is the territorialization of intelligence: the assertion that an advanced cognitive capability, however dispersed its users or its infrastructure, falls within the jurisdiction of a single state. The object of control is no longer the machine, nor even the code, but the intelligence itself. That may prove to be the most consequential legal development in the history of artificial intelligence.
This is where the Westphalian frame earns its keep. Sovereignty has always been less about territory than about the right to decide who may use a capability within and beyond one’s jurisdiction. States came to regulate cryptography, satellites, and nuclear technology not because of what those things were made of but because they redistributed power.
Frontier AI now sits in that category. The early-internet dream imagined intelligence as something that would diffuse freely across the map. The June 12 directive asserts the opposite: that advanced intelligence is a national asset, that access to it is a sovereign prerogative, and that, for the first time, citizenship may determine which intelligence a person is permitted to consult.
For most of the world, the implication is uncomfortable and specific. Governments have spent recent years pursuing compute sovereignty, building domestic data centers, securing GPU allocations, and shoring up semiconductor supply chains. The Fable episode reveals a deeper dependency they had largely ignored. A country may own its cloud infrastructure, its data, and its applications, and still lose the intelligence layer beneath all of them overnight, because that layer remains subject to a foreign jurisdiction’s authority.
For India, the Gulf states, Brazil, Indonesia, and much of Africa, the question now echoes older debates over oil and telecommunications: can essential national functions continue if access to foreign frontier models is withdrawn? India’s path differs from China’s, which has both the scale and the political will to build a parallel stack. India’s nearer objective is narrower but no less strategic, ensuring that government systems, defense applications, and its own linguistic infrastructure run on models controlled by domestic institutions.
Compute sovereignty is necessary but not sufficient. Model sovereignty is the layer that decides. More fundamentally, as artificial intelligence is territorialized, states are beginning to pursue what might be called intelligence sovereignty: the ability to generate, deploy, and govern advanced cognitive capabilities without relying on the discretionary decisions of foreign governments or firms. It is the natural endpoint of a logic that begins with chips and ends with cognition, and it will increasingly shape how middle powers think about alliances, procurement, and the architecture of their own states.
Yet the lesson of June 12 is that once a model crosses a perceived strategic threshold, the laboratory’s safety case becomes secondary to raw state interest. The decisive actor is no longer the firm but the government, acting through authorities designed for weapons and dual-use machinery. Governance is migrating from the corporate boardroom to the foreign ministry.
The arc is becoming legible. The first phase of the AI era, roughly 2022 to 2025, asked who could build the best model. The second, now opening, asks who controls access to it, through chip rules, cloud governance, and nationality-based limits. The third, later this decade, will ask who owns the intelligence infrastructure of a society, and will bring sovereign models, national AI clouds, strategic alliances, and eventually sanctions regimes built around cognition. The Fable models will likely return. The precedent will not be withdrawn. That is the signal worth recording.
The story of June 12 is not that a model went dark. It is that a major state asserted a right to control the global circulation of advanced intelligence. The internet age treated information as borderless. The AI age may treat intelligence as sovereign.
(Author is Director of Research at the Centre for Integrated and Holistic Studies, New Delhi, and formerly a Research Fellow at the Parliament of India. He writes on international law, frontier technologies and global security.)
CIHS - Centre for Integrated and Holistic Studies 