A dispute over model access: On June 12, 2026, the Commerce Department invoked federal trade regulations to bar Anthropic from distributing Fable 5 and its underlying model, Mythos 5, to foreign nationals. It was the first time the U.S. government had restricted access to a publicly available commercial AI model on national security grounds. The decision followed a warning from Amazon CEO Andy Jassy, who raised concerns after Amazon researchers extracted restricted information about cyberattacks from the Mythos model through a series of prompts. Anthropic was given mere hours to pull the model and withdraw both Fable 5 and Mythos 5 for all users globally, because the restriction extended to non-citizens working inside the United States, including some of the company’s own employees.
The dispute exposed several unresolved problems in U.S. AI policy. When does model access become a national security concern? What policy instruments apply when the object of concern is access to a digital capability rather than a physical export? How should policymakers think about the cloud platforms, GPUs, data centers, and deployment environments that make advanced AI possible? How should governance differ for open and proprietary models? And how should ordinary model usage be understood when prompting, tool use, automation, and repeated interaction can turn access into operational power or capability transfer?
These are not questions existing disciplines can answer. Advanced AI is becoming a strategic capability chain, and the frameworks built to govern software exports, national security assets, and dual-use technologies were not designed with this technology in mind. The Fable dispute provides a starting point for examining how capability becomes strategically sensitive, how access becomes a policy problem, how compute infrastructure shapes national advantage, and how AI at the frontier changes economic, scientific, and institutional resilience. Frontier AI studies is the name for the broader intellectual project required to connect those questions and develop a framework adequate to a technology whose consequences are technical, economic, institutional, and strategic at the same time.
Fable as a diagnostic event
What constitutes catastrophic capability enablement?
The primary question this emerging field needs to answer is when model access becomes a national security concern. One defensible threshold is catastrophic capability enablement: the point at which access to a model materially helps actors develop, accelerate, or operationalize cyber, biological, chemical, nuclear, or other catastrophic capabilities. That framing shifts attention from whether a model is powerful or economically significant in the abstract to what the model actually enables users to do under realistic access conditions.
The challenge is that catastrophic capability enablement is continuous in practice but still requires thresholds in governance. The same model can function as a productivity tool in one setting and a strategic asset in another, depending on the user, the domain, the tools attached to it, the scale of access, and how easily its outputs can be operationalized. Simple categories are inadequate. Decision rules, however imperfect, remain necessary. The foundational question is when model access materially moves an actor from ordinary use toward strategically dangerous action.
When does model access become a security problem?
The Fable dispute shows why model access itself has become a policy problem independent of weight release. A proprietary model that never exposes its weights can still transfer sensitive capability through APIs, cloud services, developer tools, and enterprise deployments. Users do not need the weights if they can query the model, attach tools to it, automate its outputs, or operationalize its guidance at scale.
The Commerce Department’s order addressed that exposure by restricting foreign national access, but the practical result was a full global withdrawal. The restriction extended to non-citizens working inside the United States, including some of Anthropic’s own employees. The core governance question this created is whether model access should be treated like a software service, an export-controlled technology, a national security asset, or a category the existing frameworks do not adequately describe.
When does access become capability tansfer?
Distillation sharpens the access problem considerably. Distillation is the process of training a separate model using outputs from a frontier system, allowing the capabilities of the original to be approximated without direct access to the source weights. An actor who queries a frontier model extensively can use those interactions to train, tune, or improve another model that reproduces some of the original model’s capabilities. That derived model can then operate outside the original provider’s safeguards, monitoring systems, and access controls.
Restricting weight release therefore does not fully contain the risk. Capability can migrate through repeated interaction as readily as through direct weight access. The unresolved governance question is when repeated use, prompting, automation, and distillation should be treated as ordinary model access, and when they should be treated as a pathway for transferring strategically sensitive capability to actors or jurisdictions that would otherwise be excluded.
What kind of institution can govern frontier AI?
The Fable dispute exposed a governance gap that neither government nor the private sector can close independently. Washington has legitimate authority over national security, especially where classified intelligence, foreign diversion, export controls, and adversary behavior are involved. Advanced AI capability is too technical, too fast-moving, and too dependent on deployment context to be evaluated through a purely bureaucratic process without technical depth.
The window Anthropic was given to withdraw Fable 5 illustrates the structural problem: compressed timelines, contested facts, improvised authority, and decisions made without a durable evaluative framework. A public-private institution combining national security judgment, technical evaluation, industry expertise, and proportional access controls would give both government and the private sector a more credible process for making decisions before they become emergencies. That institution would need a common vocabulary for capability, access, leakage, containment, proportionality, and innovation cost. Without it, the next Fable-level decision will be made the same way this one was.
The case for frontier AI studies
The Fable dispute shows why advanced AI needs to be studied as a strategic capability chain rather than as a set of isolated policy problems. Capability thresholds, model access, distillation risk, compute infrastructure, economic strength, labor-market resilience, and public-private governance are connected points in the same chain. Together, they determine how frontier capability emerges, becomes available, scales, transfers, and becomes strategically consequential across institutions, markets, and societies.
Frontier AI studies should become the field that examines that chain with the rigor it demands. Its purpose should be to understand when AI capability becomes a national security concern, how that capability leaks or transfers, how infrastructure shapes national advantage, and how governance can contain strategic risk without undermining U.S. leadership. The field does not yet exist in a coherent form. The Fable dispute is a signal that building it cannot wait.
The frontier AI studies agenda
Distinguishing between different kinds of model risk with greater precision is the first analytical task the field needs to complete. Open and proprietary models do not expose capability in the same way. Open models create risk through weight release, local deployment, modification, redistribution, and downstream use beyond the original developer’s control. Proprietary models retain access controls but expose capability through APIs, cloud deployment, tool integration, monitoring gaps, jurisdictional ambiguity, and distillation. These two risk profiles require different technical countermeasures and legal frameworks. The policy debate has too often conflated them.
Treating compute infrastructure as part of the AI security perimeter is the second task, and arguably the more urgent one. Advanced GPUs, accelerators, and data center capacity determine who can train, tune, distill, and deploy frontier systems at scale. Compute controls sit upstream of model controls in the capability supply chain. A country or actor that can acquire the infrastructure can build toward frontier capability even without direct access to the most advanced models. Based on ongoing research into AI infrastructure and national competitiveness, is that infrastructure access should be treated as a national security boundary alongside model access controls, not downstream of them.
Connecting AI governance to economic strength and labor-market resilience is the third task. Advanced AI affects productivity, software output, scientific progress, industrial competitiveness, occupations, skills, and the division of work between humans and machines. Those effects matter for national security because economic strength shapes the tax base, defense capacity, industrial depth, research intensity, and the ability of the United States to absorb shocks over time. A workforce that cannot adapt to AI-driven change accumulates vulnerability through displacement, weakened institutional trust, and reduced social resilience. The field needs analytical tools for weighing security risk, innovation cost, economic advantage, and labor-market adaptation within the same framework rather than treating them as separate conversations.
Beyond the dispute
Frontier AI studies should be broad enough to examine domains where model capability moves beyond general-purpose software: AI-enabled robotics, AI-accelerated biological research, autonomous cyber operations, and other settings where digital capability becomes physical, scientific, or strategic action. Those domains are extensions of the same underlying problem. The central issue is how frontier capability emerges, scales, transfers, and becomes consequential across institutions, markets, infrastructure, and society.
The category of frontier AI itself also needs scrutiny. The current policy debate focuses on models, weights, access controls, and export restrictions. Those are instruments. The larger stakes involve the political, economic, demographic, and institutional forces that will shape the next several decades. Advanced AI is becoming one of the variables acting on all of them simultaneously.
How societies organize work, distribute economic power, govern themselves, absorb demographic change, and sustain institutional trust are defining questions of the current era. Advanced AI does not sit outside those questions. It accelerates, amplifies, and in some cases destabilizes the systems through which societies have historically managed them. A field of frontier AI studies that focuses narrowly on capability thresholds, access controls, and adversarial misuse will address the proximate causes of disputes like Fable while leaving the questions that bear most directly on human welfare underexamined.
The field therefore needs to hold two levels of inquiry simultaneously: the technical and the political, economic, and social. It needs the analytical precision to evaluate capability thresholds, distillation risk, and infrastructure controls, and the scope to ask what advanced AI means for the organization of work, the distribution of economic power, and the resilience of democratic institutions. The United States has built fields of this kind before. Nuclear security, biosecurity, and space policy each required new analytical frameworks, new institutions, and new vocabularies for technologies that existing disciplines could not fully address. Frontier AI studies is the next instance of that same imperative.
The Fable dispute is the signal that the time to build it is now.
