Washington has found a way to pull frontier AI labs closer without calling it regulation. Meta is the awkward name missing from the public list.
Meta can argue about labels, openness, and whether government previews fit the way it releases AI models. It cannot argue away the direction of travel. On June 2, President Trump signed a voluntary framework that asks companies to give federal reviewers access to powerful AI models before they reach the public, and the biggest names in U.S. AI have already moved into that lane.
According to The Guardian, the order gives the government up to 30 days to review certain advanced systems before release, with national security and cybersecurity risks at the center of the exercise. The National Security Agency and the Department of Defense are part of the process, and the Commerce Department's Center for AI Standards and Innovation, or CAISI, is the office already doing much of this work inside NIST. This is not a small paperwork exercise. CAISI has completed more than 40 model assessments, including evaluations of unreleased systems, according to Commerce Department figures reported by Tom's Hardware, citing Bloomberg.
OpenAI and Anthropic had earlier evaluation arrangements with the center. Microsoft, Google DeepMind, and Elon Musk's xAI agreed in May to share models before public release. If you're building or using frontier AI, that list tells you where the pressure now sits. Washington is not waiting for Congress to pass a sweeping AI law. It is building a review habit around the companies that matter.
Meta is the obvious gap.
The company is not a side character in this story. Its Llama family made open-weight models a serious force in the market, and its AI products now sit across Facebook, Instagram, WhatsApp, and the Meta AI assistant. Business Insider reported last year that Llama 4 Scout and Maverick were released as open-weight models, while the larger Behemoth model had not been released at the time. Axios reported in April that Meta still planned to release open-source versions of upcoming models, even as some of its largest systems would probably stay proprietary. That mix makes Meta harder for Washington to handle than a closed-model lab. It also makes the absence harder to ignore.
Here is the thing: Meta's openness argument only gets you so far. Once weights are widely available, a government preview is less useful after the fact. If a model can help find software vulnerabilities, automate phishing, or lower the skill needed for cyber operations, the review has to happen before release or it becomes a postmortem with better stationery. You do not need to be hostile to open models to see the problem.
There is also a political calculation here. The June order was reportedly softened from earlier talk of mandatory preclearance, and that matters. A voluntary system lets the White House say it is protecting critical infrastructure without openly licensing AI releases. It lets companies say they are cooperating without accepting a formal regulator. Everyone gets the softer word. The harder reality is that voluntary programs can become industry standards very quickly once enough large players join.
That is why Meta's position is shrinking. A company can stand outside a voluntary framework when the framework looks experimental. It gets harder when OpenAI, Anthropic, Google, Microsoft, and xAI are already inside it, and when the government can point to dozens of completed assessments. Frankly, Meta does not want to be the company explaining why its next powerful model needed less pre-release scrutiny than everyone else's.
The business risk is not only reputational. Developers, enterprise customers, and government buyers are starting to ask different questions about model provenance, security testing, and deployment risk. A year ago, the fight was mostly about benchmark scores and model access. Now the fight includes who had the model before release, who tested it, and what kind of paper trail exists if something goes wrong.
Also read: The AI model that broke into NSA systems is now the one the NSA cannot use • Waymo's charging partner Terawatt borrows $300 million to build the physical layer beneath the robotaxi boom • Big Tech has lost $2.7 trillion in market value this month as investors question whether AI can pay for itself
None of this means the White House has solved AI safety. A 30-day preview will not catch every dangerous behavior, and a voluntary deal still leaves companies with room to define what they show and when they show it. But it changes the default. The major labs are being asked to let the government look before the rest of us touch the model.
Meta can still join and make the whole issue look temporary. If it waits, the question stops being whether Washington has overreached and becomes much simpler: why is the company with one of the most widely distributed AI strategies the last one outside the room?
