Anthropic's Export Snag: Rules No One Gets

Experts caution that current methods of AI governance, characterized by non-transparent and unsystematic interventions, are not viable long-term.

Anthropic has been working to restore its latest AI models, Fable 5 and Mythos 5, after the Trump administration issued an unexpected directive this week. This order mandated the company restrict access for all foreign nationals, a measure that ultimately led Anthropic to block access for all users, including those within the US and its own staff.

"To my knowledge, this is the first instance where US export controls have been applied to regulate access to an AI model in this specific manner."

While the Trump administration has not publicly clarified the legal foundation for its directive, Anthropic stated on its website that the government invoked "national security authorities" to justify an "export control directive" concerning its models. (Anthropic further asserted that the government's concerns regarding a potential "jailbreak" by China-linked groups to access its models were unfounded, as such methods did not bypass all of the company's protective measures.)

This raises questions about why the administration opted to utilize export control regulations for this situation. Experts describe the incident as seemingly unprecedented, highlighting an ambiguous and volatile phase in AI governance. Furthermore, the precise nature of what Anthropic is supposedly exporting remains unclear. (The company declined to comment on the matter when contacted by The Verge.)

Historically, export controls have been applied to physical items transferable across borders, such as armaments, hardware, and tools. The scope of this framework has gradually broadened to include intangible assets like software, source code, technical data, and even files for 3D-printed weapons. These items are distinct entities that can be duplicated, downloaded, published, or otherwise transferred, rather than merely accessed via a remote service like a chatbot. Regarding AI, President Joe Biden had previously sought to regulate AI model weights—the fundamental data enabling a model's function, which can be copied and deployed elsewhere—in a similar fashion; however, this approach was quickly discarded by the Trump administration during its second term.

Anthropic's situation does not align well with this established framework. No apparent transfer of goods is occurring: Mythos and Fable continue to be hosted on Anthropic's servers, and users receive only chatbot responses to their inquiries, not source code, model weights, or a copy of the model itself. While the "export" could refer to specific information generated by the models, it is unclear why this would necessitate disabling the entire system rather than simply restricting certain components. Alternatively, access itself could be considered the export, a scenario that highlights a recognized loophole in current export control regulations concerning remote access to cloud services, which Congress is actively working to address through pending Senate legislation.

Hanna Dohmen, a senior research analyst at Georgetown University’s Center for Security and Emerging Technology, informed The Verge that it remains "an open question" whether the order stretches existing regulations without access to its exact wording. She noted, "In any case, this regulation is quite notable because, to my knowledge, this is the first time US export controls have been used to control access to an AI model in this way."

"To characterize this as an unresolved domain of export control policy would be an understatement," commented Andrew Reddie, a professor at UC Berkeley’s Goldman School of Public Policy. He explained that export control rules, alongside regimes such as arms regulations, grant the government "wide latitude" to limit access to specific goods. However, he added, "the inconsistency from successive administrations regarding the responsibilities of model developers" has complicated firms' ability to comprehend their obligations.

This situation places the industry in a predicament. Should Anthropic have been targeted due to the unique capabilities of Mythos and Fable, the order prompts significant questions for upcoming models from leading AI labs such as OpenAI, Google, Meta, and xAI. If the targeting stemmed from specific safeguard deficiencies, the government must clearly define its standards for adequate protection. Furthermore, if Anthropic was singled out due to its contentious relationship with the Trump administration, the rationale behind the order becomes even more perplexing.

"This incident underscores the inherent unsustainability of the current governance framework."

Regardless of the specific reasons, experts agree that this approach to managing frontier AI is unsustainable, particularly if the US aims to retain its global leadership in the field. The incident has already bolstered arguments suggesting that international governments and companies should exercise caution when depending on American firms for access to strategically critical systems.

Reddie echoed these concerns, stating, "In some ways, I think this episode makes clear the unsustainability of the existing governance regime." He emphasized that this is particularly true if the government's primary concern was the potential for users to "jailbreak" models and circumvent safeguards. He warned, "If developing models that are impossible to jailbreak becomes the de facto standard for the United States, then it will have no AI models."

This entire situation highlights a fundamental contradiction: the Trump administration's dual stance on AI. While repeatedly advocating for a deregulated approach and supporting American technological innovation, it simultaneously compelled a leading domestic company to abruptly withdraw its advanced AI models via an order for which no public explanation has been provided. If the US government intends to regulate access to powerful AI systems, it must articulate clear guidelines and afford companies a practical opportunity to comply prior to product launch. Arbitrary, ad hoc interventions are not viable long-term and risk hindering the US's competitive position in the global AI landscape.