The Trump administration has been pressing Meta to submit its most capable AI models for federal security review, leaving the company the only major US developer that has not agreed to do so, according to a New York Times report.
The push, the paper says, has come through emails as Washington steps up oversight of frontier AI. Meta has not publicly confirmed the substance of those exchanges, and the account rests on the Times’s reporting rather than any official disclosure.
The reviews are voluntary, at least in name. They would give the government a window to evaluate a model’s abilities and weaknesses, the idea being to catch threats, from help with cyberattacks to military misuse, before a system reaches wide release.
The framework was set out in an executive order Trump signed on 2 June, which invited developers to offer “covered frontier models” to the government for up to 30 days before handing them to trusted partners.
Several of Meta’s rivals have already signed on. OpenAI and Anthropic had been working with the government on pre-release testing, and Google DeepMind, Microsoft, and xAI agreed in May to provide early access for national-security evaluations, per the Times.
That leaves Meta conspicuously outside an arrangement its peers have accepted. A spokesperson, Francis Brennan, said Meta shares the goal of advancing US leadership on “robust and secure frontier AI” and expects to sign an agreement soon, according to the report.
The timing has its own logic. Meta released Muse Spark 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 keen on early access to frontier systems would naturally turn to a developer that has just shipped a flagship model behind closed doors.
The concerns animating Washington, broadly, are that a sufficiently capable model could lower the bar for serious harm, and that the window to understand a system is narrow once it is widely deployed.
That worry is not abstract, and the administration has shown it will act on it. This month the government ordered Anthropic to bar foreign nationals from its two most capable models, a directive the company found so impractical to enforce that it switched the models off worldwide.
It showed a willingness to reach into a launched product on national-security grounds. Against that, 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 is what it is, sits awkwardly beside the scale of its AI ambitions.
The company has guided for capital expenditure well into the hundreds of billions of dollars across its AI build-out, even as it absorbs the scrutiny that has come with that spending.
A firm investing at that level has reasons to keep its newest architectures proprietary, and a review that hands the government an early look cuts against the instinct to protect what the spending bought.
What a review would actually involve is where the reporting is thinner. The Times describes the aim as identifying vulnerabilities before deployment, but the specifics of what evaluators would test, how long they would hold a model, and what they could do with the findings are governed by a framework only weeks old.
Meta’s own AI operation has had a difficult stretch, including a breach that put training secrets at risk, which complicates any talk of who gets to look inside its models.
For now the standoff is one of pressure rather than penalty. The administration has asked, Meta has said it expects to agree, and the gap between those positions is where the story sits. What is not in dispute is the direction: a government that has decided frontier AI is a security matter, and the one major developer that has not yet said yes.
Get the TNW newsletter #
Get the most important tech news in your inbox each week.