Analysis
The Trump administration has begun asserting direct control over which companies and government partners get access to frontier AI models ahead of public release, a decision that until now sat almost entirely with the labs themselves, according to CNBC sources published July 17. The shift touches Anthropic's Project Glasswing, which gave select partners early access to its Mythos cybersecurity model, and OpenAI's comparable Daybreak consortium built around GPT-5.6 -- both now face uncertainty over whether they can continue without explicit government sign-off on which partners are included.
The move follows a more direct precedent: last month the administration blocked access to Anthropic's Claude Mythos 5 and Fable 5 models over stated national-security concerns, then reinstated access only after weeks of intense negotiation with the company. That episode is now widely read inside the industry as a preview of the leverage the White House intends to hold over release timing and partner selection going forward, rather than a one-off intervention.
A White House official pushed back on the framing, telling CNBC the administration doesn't provide formal approvals for private-company AI releases, and that any government engagement, testing or meetings are 'voluntary,' with 'decisions on timing and scope of releases' resting entirely with the companies. Sources close to the labs describe the same interactions in starkly different terms -- as functionally mandatory gatekeeping that determines whether a model or partner program can proceed at all.
The dispute lands the same week Anthropic is separately negotiating a $10 billion compute lease with Meta and facing public criticism from Microsoft's Satya Nadella over its own access-control policies -- meaning Anthropic is fighting a three-front battle over who controls model access: the government above it, a hyperscaler partner beside it, and its own governance philosophy underneath it.
For enterprise AI buyers and government contractors, the practical effect is a new layer of political risk sitting on top of every frontier-model procurement decision: a model or partner program that's approved today could face a national-security hold tomorrow, on a timeline the labs themselves can't fully predict or control.
The bear case: the White House's public denial of formal approval authority means the actual legal and regulatory basis for this control remains ambiguous, and no formal rule or executive order has yet codified the practice -- it could prove to be informal pressure rather than a durable governance regime. What to watch next: whether Congress or the labs push for a formal framework governing government access decisions, and whether Project Glasswing and Daybreak survive in their current form.