| Another tantalizing question: Amazon and Anthropic’s businesses are deeply interlinked, which makes it even more significant that it was Amazon that raised the flag to the government and precipitated the intervention. (Specifically, Amazon has invested $5 billion in Anthropic and up to and $20 billion tied to commercial milestones in the companies’ partnership. Meanwhile, Anthropic has pledged to spend more than $100 billion on AWS technologies over the next 10 years.) “I’m confused why Amazon shared the jailbreak with the [U.S. government] and not Anthropic,” wrote Dave Banerjee, an associate researcher at the Institute for AI Policy and Strategy, on X. “Amazon has a large stake in Anthropic and this seems to weaken Amazon’s (financial) position.” “As a leading cloud provider that serves a large number of private and public sector customers, it’s not uncommon for governments to seek our counsel on potential security risks,” an Amazon spokesperson said in response to a request for comment. “When they occur, we don’t share the details of these discussions.” Capability & Control If the administration does make a habit of this approach going forward, the critical question becomes what statutory authority the administration is relying on to issue this export control directive. I asked the White House Sunday, and a spokesperson referred me to the Commerce Department. The Commerce Department did not respond when I followed up. According to the person who has viewed the letter, granted anonymity to describe the contents, the letter cited federal regulation section 734.13(a)(1) — which simply refers to the definition of exports under the Commerce Department’s Bureau of Industry and Security regulations. The precise legal interpretation that Commerce relies on would have broad implications for the AI ecosystem. So, I called up Joe Khawam, a managing director of legal and AI policy at the Law Reform Institute and an expert on AI-related export controls, to discuss the theories that Commerce could be relying on. Khawam outlines four possible theories: - The output theory: AI models are capable of generating information that is controlled under the Commerce Department’s Export Administration Regulations (EAR) — for instance, technical information providing blueprints for a range of topics from cyberintrusion software to nuclear technologies.
“The reason I think [the output theory] is the most likely is because it’s the broadest one out of them,” Khawam said. “Cyber software would be treated in most cases — where it doesn’t have a specific defense application as dual use — and I think would be likely controlled under the EAR.” 2. The U.S. person support theory: This theory relies on a different section of Commerce’s regulations, where the department restricts a service that it believes is being used by a foreign military intelligence or someone trying to develop a weapon of mass destruction. This “catchall” theory can encompass items that are not specifically listed in Commerce’s regulations. Rather, the theory relies on evidence regarding who is using the technology and what they are using the technology for. BIS referenced this authority in an AI policy statement last year on AI model training. 3. The weight theory: This would mirror the architecture used under the Biden administration “AI diffusion rule” that the Trump administration announced it would revoke. This framework would restrict export based on the “weights” comprising the advanced model itself. But it’s unlikely to apply here because foreign nationals are just accessing Anthropic’s models, not the underlying parameters themselves. 4. The model theory: Here, Commerce would regulate AI models themselves rather than their outputs. But it’s unlikely because BIS has “never really tried to control the model itself, as opposed to the weights or something else that can fit within their definitions of technology or software,” said Khawam. It’s an idea that, at least rhetorically, administration officials have been ideologically opposed to from the get-go. | “Each of [the arguments] has vulnerabilities that make it somewhat difficult to really rely upon as a legal basis to pull down the entire system… [With the most likely output-based theory] and you say take the entire model down, because it has the potential to generate controlled outputs, then there is no actual export that you can point to that has occurred… It’s more like a red flag type directive, which again I don’t think it’s beyond the pale that they could make that argument in a directive issued to Anthropic. I think it’s just an aggressive use of their authorities.” Joe Khawam, a managing director of legal and AI policy at the Law Reform Institute | | | | The implications: The direction that the government takes with an export control regime would have massive implications for different parts of the U.S. AI industry. Under the output theory, AI models become regulated based on what they can say. But that creates enormous uncertainties for the labs since every model can generate some level of cyber, chemistry and nuclear information that could hypothetically be used in a nefarious way. The second theory potentially gives the U.S. government wide discretion. If the administration says it has intelligence indicating foreign military interest in an AI model, the government can then move to shut it down. The third theory would spell big trouble for open-source models, such as those pioneered by Meta, French AI company Mistral and the U.S.-based Reflection AI. Publishing the weights of AI models is at the heart of these open-source companies, so they would be effectively hamstrung if the administration decided to go in this direction. What’s happening now: Cybersecurity specialists have been signing on to an open letter asking Commerce Secretary Howard Lutnick and National Cyber Director Sean Cairncross to lift the export restrictions on Anthropic’s Fable and Mythos models. “This action has taken the best models away from defenders, created market uncertainty, and risked America’s AI leadership without any real risk to justify it,” they wrote in the open letter. Signers include Corridor Chief Product Officer Alex Stamos and Aaron Brown, head of security at Mercor. This newsletter is published by WP Intelligence, The Washington Post’s subscription service for professionals that provides business, policy and thought leaders with actionable insights. WP Intelligence operates independently from The Washington Post newsroom. Learn more about WP Intelligence. | 
