Anthropic Faces Confusion Over New Export Regulations
Anthropic has been scrambling this week to restore access to its latest AI models after an unexpected directive from the Trump administration mandated that the company restrict access for all foreign nationals, including users in the US and its own staff. This led to the blocking of Fable 5 and Mythos 5 for everyone.
While the Trump administration hasn’t clarified the legal grounds for this order, Anthropic stated that the government referenced “national security authorities” to support an “export control directive” concerning the models. The company also noted that the government’s worries about a potential “jailbreak” that could allow groups linked to China to exploit its models did not mean users could bypass all of Anthropic’s security measures.
The use of export control rules in this context raises questions. Experts suggest this situation is unprecedented, highlighting the current instability in AI governance. What exactly is Anthropic meant to be exporting? The company did not provide a comment on the matter.
Traditionally, export controls apply to tangible items that can be shipped across borders, like weapons and hardware. Over time, this has expanded to include less tangible assets such as software and technical data. However, these are still discrete items that can be copied or transferred, unlike services accessed remotely, such as chatbots. In the realm of AI, President Biden had previously sought to regulate AI model weights—the essential data that enables a model to function—though this initiative was quickly abandoned by the Trump administration.
The directive concerning Anthropic doesn’t fit neatly into existing frameworks. There’s no clear transfer of assets occurring: Mythos and Fable are hosted on Anthropic’s servers, and users don’t receive source code or model weights; they only get responses from the chatbot. The export could refer to specific outputs generated by the models, but it’s unclear why this would necessitate cutting off access entirely rather than just limiting certain aspects. It might also pertain to access itself, although remote access to cloud services is a known gap in current export control laws, which Congress is attempting to address with new legislation.
Hanna Dohmen, a senior research analyst at Georgetown University’s Center for Security and Emerging Technology, remarked that it remains “an open question” whether the order strains existing regulations without knowing the exact language behind it. She noted that this is likely the first instance of US export controls being applied to restrict access to an AI model in this manner.
Andrew Reddie, a professor at UC Berkeley’s Goldman School of Public Policy, emphasized that the current state of export control rule-making is quite unsettled. He pointed out that while export control regulations grant the government considerable authority to limit access to certain goods, the inconsistent messaging from various administrations regarding the responsibilities of model developers has left companies uncertain about what is expected of them.
This situation puts the industry in a difficult position. If Anthropic was targeted due to the unique capabilities of Mythos and Fable, it raises significant questions for future models from other leading labs like OpenAI, Google, Meta, and xAI. If the focus was on specific safeguard issues, the government needs to clarify what protections it deems adequate. If Anthropic was singled out due to its contentious relationship with the Trump administration, the rationale becomes even murkier.
Experts agree that this approach is not a viable way to manage advanced AI, especially if the US aims to maintain its global leadership. The incident has already fueled concerns that governments and companies outside the US should be cautious about depending on American firms for access to critical systems.
Reddie echoed these concerns, stating that this episode underscores the unsustainability of the current governance framework. This is particularly true if the government is more worried about users being able to jailbreak models and bypass safeguards. “If creating models that are impossible to jailbreak becomes the de facto standard for the United States, then it will have no AI models.”
Ultimately, this situation illustrates a fundamental issue: the Trump administration is trying to have it both ways with AI. It has repeatedly expressed a desire for a hands-off approach while promoting American technology, yet it has compelled a domestic leader to abruptly withdraw its advanced models through an unexplained order. If Washington intends to regulate access to powerful AI systems, it must clarify its methods and provide companies with a genuine opportunity to comply before launching. Arbitrary interventions are not a sustainable strategy and could jeopardize the US’s position in the AI race.
