Every major US artificial intelligence developer has signed an agreement to voluntarily share its models with the federal government for safety review. Every one except Meta.

The Trump administration sent emails to Meta in recent weeks pressing the company to submit its AI models for evaluation by the Center for AI Standards and Innovation (CAISI), a unit housed within the Commerce Department, according to The New York Times, which cited four people familiar with the confidential request. The review process involves assessing a model’s capabilities and vulnerabilities, the people said, speaking on condition of anonymity because they were not authorized to discuss the matter publicly.

OpenAI, Anthropic, Google, xAI, and Microsoft have all reached voluntary agreements with CAISI. Meta, which released its most recent AI model in April 2026, has not. Its policy team has been in negotiations with the Commerce Department over how to proceed, but no deal had been finalized as of June 23, the Times reported.

Meta issued a statement through spokesman Francis Brennan acknowledging the gap. “We share the administration’s goal of advancing US leadership on robust and secure frontier AI,” Brennan said. “While we are working through the details, we hope to sign the agreement soon.” Ben Kass, a Commerce Department spokesman, described the situation as routine: “This is the very work CAISI is supposed to be doing.”

The context matters. Less than two weeks before this request became known, the administration ordered Anthropic to remove access to its newest model, citing national security grounds. That episode signals the federal government is willing to act coercively when voluntary frameworks stall. The open-ended nature of CAISI’s review authority means the administration has few formal levers to compel participation but significant informal pressure to apply.

Meta’s position as the lone holdout is structurally distinct from the rest of the field because of its open-weight strategy. OpenAI, Anthropic, and Google release through APIs and closed products where access can be revoked. Meta publishes model weights through the Llama family, meaning any model submitted for federal review could, in principle, already be running on tens of thousands of independent servers before the review concludes. Whether CAISI’s evaluation process is designed to account for that distribution reality is a question the administration has not publicly answered.

The commercial stakes for the Llama ecosystem are real. Thousands of companies build on Llama weights for fine-tuning, inference, and embedded deployment. If a review surfaces capability concerns that the government acts on, those downstream builders face disruption with little warning. A voluntary agreement that includes advance notice provisions would provide more stability than the current ambiguity.

The Times noted that some voices inside the industry argue the cybersecurity concerns driving these reviews are overstated, on the grounds that capable AI systems can defend networks as readily as they can attack them. That argument does not appear to have slowed the administration’s push.

Any team currently using Llama 4 or planning to build on the next Llama release should treat the status of this agreement as a live procurement risk variable until Meta and the Commerce Department publicly confirm a deal is in place.

Reported by Tripp Mickle, Eli Tan, and Sheera Frenkel for The New York Times on June 23, 2026.