Anthropic announced Thursday that it had finally restored limited access to its Fable 5 and Mythos 5 conversational agents after a week of disruption caused by a sudden directive from the Trump administration. The order required the company to block the models for every foreign national, a category that, under the directive, also encompassed users inside the United States and even Anthropic employees. The government’s justification referenced "national security authorities" and an "export control directive," but no detailed explanation has been released.
Company officials said the move was prompted by worries that a "jailbreak" could enable actors linked to China to bypass the models' built‑in safeguards. Anthropic’s statement stressed that the order forced it to shut down the services entirely rather than apply a more targeted restriction on specific outputs or capabilities. The abruptness of the action left the firm scrambling to re‑enable the services while seeking clarification on the legal basis for the order.
Legal scholars and policy analysts note that this is the first time U.S. export controls have been used to limit access to an AI model hosted as a remote service. Traditionally, export regulations have covered tangible items—arms, hardware, technical data—that can be physically shipped or copied. Over recent years the scope has broadened to include software source code and even 3D‑printed weapon files, but those still involve a transfer of discrete files. In Anthropic’s case, the models remain on the company’s servers, and users only receive generated text, not the underlying weights or code.
President Biden earlier in his term proposed restricting AI model weights, the data that enables a model to be reproduced elsewhere, but his administration abandoned that approach. The current directive, issued under the Trump administration, does not fit neatly into existing frameworks because there is no obvious export of the model itself. Some experts suggest the export could be the specific information produced by the AI, yet the order disables the entire system rather than isolating particular outputs.
"It is an open question whether the order strains existing rules without seeing the precise language behind it," said Hanna Dohmen, senior research analyst at Georgetown University's Center for Security and Emerging Technology. Andrew Reddie, a professor at UC Berkeley's Goldman School of Public Policy, warned that while export‑control regimes give the government broad latitude, the lack of clear guidance makes compliance difficult for firms developing frontier AI.
The episode puts the entire industry in a bind. If Anthropic was singled out because its models are uniquely powerful, the move could set a precedent that affects OpenAI, Google, Meta, xAI and other labs developing next‑generation systems. If the concern centers on safeguard shortcomings, regulators will need to articulate concrete expectations. And if the action reflects a strained relationship between the company and the administration, it underscores the unpredictability of policy enforcement.
Experts argue that the current patchwork approach to AI governance is unsustainable. Reddie cautioned that demanding models be "impossible to jailbreak" as a de‑facto standard could leave the United States without any viable frontier AI offerings. The lack of a transparent, consistent framework may also push foreign governments and companies to seek alternatives outside the U.S., potentially eroding America’s leadership in the field.
Congress is already considering legislation to close gaps in export‑control coverage of cloud‑based services, but until such laws pass, companies like Anthropic must navigate an uncertain regulatory landscape. The Anthropic incident highlights the tension between national‑security objectives and the rapid pace of AI innovation, raising the question of whether ad‑hoc interventions can ever keep pace with the technology they aim to regulate.
Questo articolo è stato scritto con l'assistenza dell'IA.
News Factory APP - notizie agentiche per potenziare il tuo SEO e AEO.