Every Major AI Lab Agreed to US Safety Review — Except Meta

The US Center for AI Standards and Innovation — an evaluation body operating under Commerce Secretary Howard Lutnick — has asked Meta to submit its Muse Spark AI model for government review before it goes public. The request follows an executive order signed on June 2 that establishes a formal framework for evaluating frontier AI models: officials get up to 30 days to assess a model's capabilities and identify potential security vulnerabilities before any public release.

The striking part: Meta is the only major AI lab that has not yet agreed to this arrangement. OpenAI, Anthropic, Google, xAI, and Microsoft have all provided early access to their models for government review. Meta spokesperson Francis Brennan responded to the government's request by saying the company "shares the administration's goal of advancing US leadership on robust and secure frontier AI" — without confirming whether Meta would actually comply.

How the new review framework works

Under this setup, AI companies hand over access to their most capable models before those models are released to the public. The review window is up to 30 days, and the evaluation focuses on what the model can do, where it might be misused, and whether any capabilities warrant additional safeguards or restrictions. Think of it as a brief inspection period between a model being ready and it becoming publicly available — similar in spirit to how new drugs or devices are reviewed before they can be sold.

Meta has historically taken a more open stance than its competitors: it open-sourced its LLaMA and Muse model families instead of keeping them behind closed, paid APIs. That open-release approach may partly explain the friction here. Submitting to a pre-release government review sits awkwardly alongside a culture of releasing models directly to the public — you can't simultaneously open-release a model to everyone and hand it to the government for a 30-day exclusive preview.

Why AI builders should pay attention

This is the first formal pre-release review framework for AI models in the United States, and five of the six largest AI developers are already operating inside it. If Meta eventually joins — and the pressure to do so is real — this creates a new checkpoint in the AI development process that didn't exist a year ago: before a frontier model reaches developers and users, it passes through a government review window.

For people building products on AI APIs, the short-term practical impact is minimal: models still ship, just with a government review window added to the timeline. But the longer arc is significant. This is the beginning of a formalized relationship between frontier AI development and US government oversight, with some structural parallels to how high-stakes products in other industries get reviewed. The question of whether this improves safety, slows deployment, or changes what models get built in the first place is genuinely open.

Muse Spark — Meta's most capable current model, featuring both standard and extended reasoning capabilities — is the specific subject of the current review request. The model launched in April; the government's request suggests officials want a look at what Meta's reasoning-capable model can actually do before it gets wider distribution.