The White House is now determining who has access to advanced AI models, rather than the laboratories.

      TL;DR: The Trump administration is controlling access to advanced AI models, transferring that authority from Anthropic and OpenAI to the government via the Gold Eagle program.

      According to a CNBC report on Friday, the Trump administration is now determining which companies and organizations can access advanced AI models from Anthropic and OpenAI, citing two sources familiar with the situation. Previously, these labs had full autonomy over those decisions. Anthropic managed access to its Mythos cybersecurity model through an initiative called Project Glasswing, while OpenAI operated a similar program named Daybreak for its cyber model. Moving forward, partner lists will require explicit approval from the government.

      A White House representative mentioned to CNBC that the government does not grant "approvals for AI releases" and that participation from companies is "voluntary." However, last month, the administration halted access to Anthropic’s Claude Mythos 5 and Fable 5 over national security issues, only restoring it after lengthy discussions. In June, OpenAI announced it would restrict new models to "trusted partners" to align with government directives. The discrepancy between the official stance and the practical situation is noteworthy. This week, the White House introduced Gold Eagle, an AI clearinghouse for cyber vulnerabilities, and according to a source from CNBC, the program will enable the White House to authorize which companies can access new AI models.

      The timing raises political concerns. Moonshot AI launched its Kimi K3 on the same day, achieving performance comparable to or exceeding Fable and GPT-5.6 on at least one independent benchmark. David Sacks, the former White House AI czar, deemed this "concerning," stating: "This is how you lose the AI race. The rest of the world won’t adhere to our rules if we hinder ourselves." The administration aims to protect frontier AI from Chinese use while simultaneously observing Chinese laboratories catch up in real time.

      This change is structural rather than temporary. An executive order from Trump in June requested AI companies to provide the government with early access to models for testing, presented as voluntary. Gold Eagle transforms that request into something resembling a gating mechanism. If Anthropic and OpenAI are unable to release their most advanced models without government approval of their partner list, the US government effectively gains distribution authority over frontier AI, without the need for legislation, a regulatory body, and through a program the White House claims is optional.