When governments think about controlling the spread of dangerous technology, they typically imagine physical objects: missiles, centrifuges, specialized chips. AI models present a fundamentally different problem. A piece of software capable of advanced reasoning can cross borders in seconds, replicated perfectly with no degradation, making the traditional tools of export control look poorly suited for the task. A new analysis from the Bulletin of the Atomic Scientists takes a close look at why models like Claude Fable and Mythos are forcing that reckoning.
The Problem With Controlling Code
Export control regimes were designed around scarcity. A nation that controls access to enriched uranium or precision machine tools can meaningfully slow a rival's weapons program. AI model weights do not work that way. They can be copied, distributed, and run on hardware that is increasingly available worldwide. When Anthropic disabled Fable 5 and Mythos 5 following a U.S. export order, the action raised an immediate question: what exactly had been controlled, and for how long could that control hold?
Key Facts
- Traditional export controls target physical goods with inherent scarcity; AI model weights can be copied infinitely at near-zero cost.
- The Bulletin of the Atomic Scientists argues current frameworks lack technical definitions adequate for frontier AI systems.
- Claude Fable and Mythos represent a new capability tier that regulators did not anticipate when existing rules were written.
- Compliance mechanisms for software exports remain largely unverified once a model is deployed abroad.
- Experts warn that unilateral controls may disadvantage U.S. companies without meaningfully slowing adversary access.
The Bulletin's analysis zeroes in on a structural gap: export control categories were built around performance thresholds for hardware, not for the emergent capabilities of large language models. Defining when a model becomes export-restricted requires agreeing on what makes it dangerous, and that agreement does not yet exist at a regulatory level. Anthropic's IPO filing, submitted just weeks before the government shut down access to its most capable models, listed export restrictions as a material risk factor, signaling that the company itself views regulatory exposure as a serious business variable.
"The challenge with AI models is that the technology is the information, and information controls have always been the hardest to enforce."Bulletin of the Atomic Scientists analysis, 2025
What Fable and Mythos Represent
The specific models named in the Bulletin's piece are not incidental. Claude Fable and Mythos occupy a capability class that sits above standard commercial AI assistants. Since Anthropic launched Claude Fable as its first Mythos-class model, policy observers have debated whether existing dual-use technology rules apply, and if so, how. The models can perform complex scientific reasoning, assist with code at a professional level, and process information in ways that may be relevant to sensitive domains. That profile is exactly what export control architects worry about, but the statutes they work with were not written with such systems in mind.
Part of what makes enforcement difficult is verification. When a company ships a controlled physical component, customs officials can inspect it. When a model is accessed via an API or downloaded as weights, the compliance chain becomes harder to audit. Regulators can issue orders, as they did with Fable and Mythos, but monitoring adherence across global user bases is a different challenge entirely. The Bulletin's authors argue that without new technical and legal mechanisms, export orders may function more as liability shields for companies than as genuine barriers to proliferation.
A Framework Built for a Different Era
The deeper issue the Bulletin raises is one of institutional lag. The agencies responsible for export controls, and the statutes they administer, evolved over decades in response to nuclear and missile technology. Adapting those frameworks to cover AI requires not just new regulations but new technical definitions, new enforcement infrastructure, and international coordination that does not currently exist. In the meantime, frontier AI models continue to advance. Anthropic and its peers are developing systems faster than policy can respond, a dynamic that is familiar from other technology domains but carries sharper stakes when the technology in question can assist with scientific research at the cutting edge.
The debate is unlikely to resolve quickly. Export control reform requires legislative action, agency rulemaking, and diplomatic engagement, all moving on timelines that bear little resemblance to AI development cycles. What the Bulletin's analysis makes clear is that Claude Fable and Mythos are not just products caught in a regulatory dispute. They are test cases for whether existing legal infrastructure can handle a class of technology it was never designed to govern.
