The Trump administration has been pressuring Meta to submit its most capable AI models for a federal security review, making the company the only major US developer that has not agreed to do so, according to a report. New York Times report.
The push, the paper says, came via emails as Washington steps up oversight of border AI. Meta has not publicly confirmed the substance of those exchanges, and the account is based on reporting by the Times rather than any official disclosures.
Reviews are voluntary, at least nominally. They would give the government a window to evaluate a model’s capabilities and weaknesses, with the idea of detecting threats, from help with cyberattacks to military misuse, before a system reaches widespread release.
The framework was laid out in an executive order Trump signed on June 2, which invited developers to offer “covered border models” to the government for up to 30 days before turning them over to trusted partners.
The 💜 of EU technology
The latest rumors from the EU tech scene, a story from our wise founder Boris and some questionable AI art. It’s free, every week, in your inbox. Register now!
Several of Meta’s rivals have already signed up. OpenAI and Anthropic had been working with the government on pre-launch testing, and Google DeepMind, Microsoft and xAI agreed in May to provide early access for national security assessments, according to the Times.
That leaves Meta visibly outside a deal its peers have agreed to. A spokesman, Francis Brennan, said Meta shares the goal of promoting American leadership in “Robust and secure border AI” and hopes to sign a deal soon, according to the report.
The moment has its own logic. Goal released spark muse in April, the first model from its Superintelligence Labs unit and, in a break from the company’s Llama heritage, a closed one.
A government interested in early access to border systems would naturally turn to a developer that has just launched a flagship model behind closed doors.
The concerns animating Washington, broadly speaking, are that a sufficiently capable model could lower the bar for serious harms and that the window for understanding a system is narrow once it has been widely deployed.
That concern is not abstract, and the administration has demonstrated that it will act accordingly. This month, the government ordered Anthropic to bar foreign nationals from accessing its two most capable models, a directive the company found so impractical to enforce that turned off models around the world.
Showed willingness to access a product launched for national security reasons. In contrast, an invitation to submit models for review reads less as a courtesy than as the softer end of a spectrum that has a harder end.
Meta’s reluctance, if that’s what it is, sits uncomfortably alongside the scale of his AI ambitions.
The company has planned capital expenditures of up to hundreds of billions of dollars across its AI development, even as it absorbs the scrutiny that has arrived with that expense.
A company that invests at that level has reasons to keep its newer architectures proprietary, and a review that gives the government an early look goes against the instinct to protect what was bought with the expense.
What a review would really imply is that the reports be more scarce. The Times describes the goal as identifying vulnerabilities before deployment, but the details of what testers would test, how long they would maintain a model and what they could do with the findings are governed by a framework just weeks old.
Meta’s own AI operation has had a rough patch, including a Violation that puts training secrets at risk.which complicates any conversation about who can look inside their models.
For now, the confrontation is more about pressure than penalties. The administration has asked, Meta has said it hopes to agree, and the gap between those positions is where the story lies.
What’s not up for discussion is direction: a government that has decided that cutting-edge AI is a security issue, and the only major developer that hasn’t yet said yes.
