Meta spent the past two years selling itself as the open alternative in American AI, the company that gave Llama away while its rivals locked their best work behind paywalled interfaces. This week it found itself isolated for the opposite reason. According to a New York Times report, citing four people familiar with the confidential request, Meta is now the only major US AI developer that has not agreed to submit its most capable models to the federal government for security review before they ship.
The mechanism is new and still half-built. On June 2, President Trump signed an executive order setting up a voluntary framework under which developers offer "covered frontier models" to the government for up to 30 days before handing them to trusted partners. The point is to catch the worst dangers, help with cyberattacks or military misuse, while the window to understand a fresh system is still open. The Commerce Department's Center for AI Standards and Innovation, under Secretary Howard Lutnick, runs the evaluations, and has until the end of July to finalise the process. OpenAI and Anthropic were already testing unreleased models with officials; Google DeepMind, Microsoft and xAI agreed in May to grant early access.
That leaves Meta conspicuously outside an arrangement its peers have accepted. The requests arrived by email, and spokesperson Francis Brennan offered a careful non-denial: "We share the administration's goal of advancing US leadership on robust and secure frontier AI. While we are working through the details, we hope to sign the agreement soon." The reluctance has a logic. In April Meta shipped Muse Spark, the first model from its Superintelligence Labs and, in a break from its open heritage, a closed one. A company that has guided to hundreds of billions in AI spending has every reason to guard the architectures that money bought, and a reported breach that put training secrets at risk makes the question of who gets to look inside even touchier.
What gives the word "voluntary" its weight is a precedent only weeks old. This month the government ordered Anthropic to bar all foreign nationals from its two most capable models, the cybersecurity system Mythos 5 and its public sibling Fable 5. Anthropic found the directive so impractical to enforce that it switched both models off worldwide rather than risk breaching it. Set against that, an invitation to submit a model for review reads less as a courtesy than as the soft end of a spectrum whose hard end is a global kill switch. The administration has shown it will reach into a launched product when it decides national security is at stake.
Step back and the Meta standoff is a single frame in a longer movement. US export controls began with physical chips, then expanded to the equipment that makes them, then to cloud access that routes around hardware entirely, and now to frontier models themselves; Congress is weighing the Chip Security Act, the MATCH Act, the Remote Access Security Act and the AI OVERWATCH Act in parallel. The throughline is a decision, now close to settled, that a frontier model is a strategic artifact to be inspected before release, like a munition or an advanced chip. Meta's holdout is not really about Meta. It is the moment the model layer itself became something governments expect to vet, and the only open question is how loudly the last holdout signs on.