U.S. Presses Meta to Join the Government’s AI Safety Review

The Trump administration is pressing Meta to submit its artificial intelligence models for voluntary federal review, the only major U.S. AI developer that has not agreed to the program. The request was made by email, four people familiar with the matter told The New York Times. Meta’s absence is the most visible gap in a federal voluntary review program the other five major U.S. AI developers have already joined.

Meta spokesman Francis Brennan said the company is “working through the details” and hopes to sign soon. The Commerce Department, which houses the review program, said the outreach is part of its routine work. The pressure campaign comes less than two weeks after the same government ordered Anthropic to disable two of its newest models for foreign users, citing national security concerns.

The Government Wants Meta Inside the Voluntary Review

Federal officials have been emailing Meta in recent days, asking the company to open its frontier AI models to government testers before public release, four people familiar with the confidential request said. The reviews are run by the Center for AI Standards and Innovation, or CAISI, a Commerce Department body that evaluates whether advanced AI systems pose cybersecurity, biosecurity or chemical weapons risks. The arrangement is the same one the other major U.S. AI developers have already accepted, with Meta the only one that has not reached an agreement to voluntarily share its models with the federal government for review.

Meta’s response, when it came, was noncommittal. The company framed itself as wanting to cooperate, but stopped short of committing. The Commerce Department, for its part, framed the outreach as routine work. Both sides stopped short of disclosing the specific terms still in dispute.

We share the administration’s goal of advancing U.S. leadership on robust and secure frontier A.I. While we are working through the details, we hope to sign the agreement soon.

The statement came from Francis Brennan, a Meta spokesman, on Tuesday. The Commerce Department’s response was more clipped. “This story is not unusual,” spokesman Ben Kass said. “It is the very work CAISI is supposed to be doing.” The two responses, taken together, capture the gap: the White House is treating Meta’s absence as a problem to solve, while the agency treats it as a routine conversation.

Five Labs Have Signed

The roster of companies that have agreed to voluntary pre-release CAISI review has grown steadily over the past two years. OpenAI and Anthropic were the first to sign comparable agreements, in 2024, when the precursor body was still called the U.S. AI Safety Institute. Google DeepMind, Microsoft and xAI followed in May 2026, according to the program tracker AI Weekly. Meta is the only major U.S. AI developer that has not yet joined.

For each lab, the deal is structured the same way. The government gets a window of access to new models before they ship to other partners. Testers run them through classified and unclassified evaluations, looking for the kinds of cyber or biosecurity capabilities that would worry national security agencies, and the findings feed back to the company to inform any “covered frontier model” designation under the Trump administration’s June 2 executive order.

What Just Happened to Anthropic

The Trump administration’s new pressure on Meta follows a direct action against another frontier lab. On June 12, 2026, the government issued an export control directive ordering Anthropic to suspend all access to its Fable 5 and Mythos 5 models by any foreign national, inside or outside the United States. The directive reached the company at 5:21 p.m. Eastern Time the same day, and Anthropic said the order applied to its own foreign national staff as well as outside customers. The practical effect is a hard shutdown, with the two models no longer accessible to anyone.

Anthropic’s understanding, laid out in a public statement that day, was that the government had become aware of a “jailbreaking” method, a way to bypass the model’s safety guardrails, that could expose specific software vulnerabilities. The government provided only verbal evidence, the company said, and Anthropic reviewed a report on which it believed the directive was based. The company concluded that the level of capability on display was “widely available from other models” and “used every day by the defenders who keep systems safe.”

The result is a hard shutdown. The two models are no longer accessible to any Anthropic customer, foreign or domestic, including Anthropic’s own foreign national staff. Anthropic said the order is technically an export control measure, not a model recall, though the practical effect is the same. The directive was not issued under the CAISI voluntary review process, and the two are distinct mechanisms.

Anthropic complied. The company also pushed back. “We are complying with the government’s legal directive and are removing access to Fable 5 and Mythos 5 for all users,” the Anthropic statement on the suspension said. “However, we disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.” Anthropic’s argument is that, if applied industry-wide, the standard would “essentially halt all new model deployments for all frontier model providers.”

The Anthropic directive is not a CAISI voluntary review, and the two should not be confused. It is the kind of unilateral national security action the government can take on its own, separate from any voluntary pact. For a company weighing whether to sign the CAISI voluntary agreement, however, the case study is now sitting in the public record. The voluntary framework promises a longer runway and a feedback loop, while the alternative, as Anthropic just learned, can be a single email that disables a model for everyone.

The Framework Behind the Request

The voluntary review program Meta is being asked to join was formalized three weeks before the request. On June 2, 2026, President Trump signed Executive Order 14409 on AI innovation and security, which set a uniform federal structure for engaging frontier AI developers. The order shortened the proposed pre-release access window from an initial 90 days to up to 30 days and made clear that participation is not mandatory.

The order explicitly says nothing in it “shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models.” That language is what gives the program its voluntary character. It is also what makes Meta’s absence a problem of persuasion rather than enforcement. The administration has chosen the softer tool, at least for now.

CAISI sits at the center. Housed at the National Institute of Standards and Technology, the Center for AI Standards and Innovation “will serve as industry’s primary point of contact within the U.S. government to facilitate testing and collaborative research” on commercial AI systems, per its published mission and voluntary review mandate. It “establish[es] voluntary agreements with private sector AI developers and evaluators” and runs evaluations aimed at “demonstrable risks, such as cybersecurity, biosecurity, and chemical weapons,” and operates a classified benchmarking process to determine when a model crosses the line into a “covered frontier model” designation that triggers federal pre-release access.

1. August 2024: OpenAI and Anthropic sign first voluntary agreements with the U.S. AI Safety Institute, the body that later became CAISI.
2. May 2026: Google DeepMind, Microsoft and xAI reach CAISI pre-release evaluation agreements, bringing the participating lab count to five.
3. June 2, 2026: President Trump signs Executive Order 14409, formalizing a voluntary 30-day pre-release review window and codifying CAISI’s role.
4. June 12, 2026: The government orders Anthropic to suspend Fable 5 and Mythos 5 access for foreign nationals, citing national security concerns.
5. June 23, 2026: Federal officials email Meta, asking the company to agree to voluntary CAISI review of its models.

Why Meta’s Absence Stands Out

Five frontier labs in, and the missing chair is the one most Americans would recognize. Meta is the only major U.S. AI developer that has not signed a CAISI voluntary review agreement. It has built its recent AI messaging around open and rapid release, and its most recent frontier model shipped in April 2026 without the kind of pre-release government testing every other major peer accepted. The asymmetry is what makes the request news.

The Trump administration, which spent its first year rolling back Biden-era AI safety measures, is now investing political capital in getting one specific company to the table. A voluntary framework that every major peer has joined leaves Meta with limited room to argue the program is unfair or experimental. The administration is also publicly considering whether to make some form of CAISI review mandatory in the future, a step that would change the calculation for any holdout.

Meta’s calculus is straightforward. Voluntary review means ceding some early access to government testers, including access in classified settings. The trade gets more expensive the longer Meta is the only one making it. And the alternative, after the Anthropic directive, looks more like a unilateral national security action than a negotiated agreement.

Whether the program itself holds up is a separate question. Devin Lynch, a former director for cyber policy at the White House Office of the National Cyber Director, wrote on LinkedIn that “CAISI will need to define, and publish, what it’s testing for, not just who it’s testing with.” The reviews only have value if the threat models behind them are credible, and the practical question is whether the program can pull Meta to the table without any compulsory power behind it.

Voluntary Review in Practice

The voluntary label cuts both ways, and that is by design. OpenAI, Anthropic, Google, Microsoft and xAI agreed in part because the program carries no legal compulsion, and the participating labs keep control over release timing. Meta can decline without an immediate penalty for the same reason. The Trump administration has so far stuck to the line that the program is voluntary, even as it ramps up pressure on Meta to participate.

The pre-release window is narrow. Under the order, covered frontier model developers can give the government access for a period of up to 30 days before they plan to release such models to other trusted partners. OpenAI’s published work with CAISI on security has produced concrete fixes to its products within a business day, a model the agency points to as evidence the program delivers value beyond paper compliance.

How Meta Will Respond

Meta’s stated position is that it wants in. The Brennan statement, soft as it is, signals a company that has not dug in against the program. The language points to a firm that expects to be at the table eventually, on terms it has not yet defined. Meta has not said when it would sign, only that it hopes to do so soon.

The other side of the same posture is the one the White House is responding to. CAISI’s job, as Ben Kass put it, is to “regularly engage[] with companies about voluntary agreements,” and Meta is the one major U.S. AI developer that has not engaged on the federal review. The administration’s separate consideration of mandatory reviews, on top of the recent directive against another frontier lab, gives the voluntary ask a sharper edge than its language suggests.

Meta’s next model release will be the first public test of the company’s stated willingness to sign. Whatever Meta ships, the result moves the question from press leaks to a model anyone can look at and evaluate against the others. The program gets its first visible answer either way. Brennan’s “soon” is the only public timeline the White House has to work with.