The U.S. government and AI company Anthropic are negotiating the return of powerful AI models that were recently taken offline due to national security concerns.
This situation began when Anthropic was forced to disable its new frontier models, Fable 5 and Mythos 5, just days after their launch. The direct cause was a U.S. export-control order aimed at preventing access by foreign nationals. The government was particularly concerned about jailbreaking risks—users tricking the AI into bypassing its safety features—and the potential for the technology to be misused. An urgent alert from Amazon Web Services (AWS), Anthropic's cloud partner, reportedly highlighted these risks and accelerated the government's decision to order the shutdown.
Initially, the shutdown seemed like a definitive ban, especially given the Trump administration's previously hostile stance, which included an order in February for government agencies to stop using Anthropic's technology. However, the narrative shifted from a permanent ban to a negotiation. President Trump's recent comments, stating that talks are "going fine" and relations have "improved," signaled a willingness to find a solution. This transformed the shutdown from a final verdict into a powerful bargaining chip to bring Anthropic to the table under the government's terms.
Now, the talks are centered on establishing a new governance framework described as "control-and-verify." The likely path forward is guided by a recent executive order that created a voluntary channel for AI companies to submit their models for White House cybersecurity testing before release. The goal is to allow a controlled, U.S.-only relaunch of the models with much stronger safeguards, such as enhanced identity verification, geo-restrictions, and continuous monitoring. The outcome of these negotiations is significant because it will likely set a precedent for how the U.S. regulates all high-capability AI, influencing major tech players from cloud providers like AWS to defense contractors like Palantir.
- Jailbreaking: A technique used to bypass an AI model's built-in safety rules and restrictions, often to generate prohibited or harmful content.
- Export Control: Government regulations that restrict the transfer of certain goods, technologies, and software to foreign countries or foreign nationals for reasons of national security or foreign policy.