TL;DR
· According to media reports, the Trump administration is pushing Meta to submit its cutting-edge AI models for voluntary review by the federal government.
· EO 14409 requires the design of a voluntary framework within 60 days, allowing developers to provide up to 30 days of access before release.
· This mechanism is not a mandatory approval, but incidents involving Anthropic show that model access may be influenced by safety judgments.

According to media reports, the Trump administration is urging Meta to submit its frontier AI models for voluntary review by the federal government, with Meta being described as one of the few major U.S. AI developers not yet to have signed such an arrangement. For AI companies, this is not a traditional approval process, but it could affect how the most powerful models are tested for safety before public release, what capabilities the government can see in advance, and whether access might be restricted if national security risks involving cyber, biological, or chemical weapons are involved. Companies like OpenAI, Anthropic, Google, xAI, and Microsoft are reportedly agreeing to cooperate with agencies under the Department of Commerce, which has amplified the progress of Meta’s negotiations.

This engagement mainly revolves around the U.S. Artificial Intelligence Standards and Innovation Center (CAISI). CAISI, under the Department of Commerce and NIST, was reorganized in June 2025 from the U.S. AI Safety Institute established during Biden’s administration, with a focus on voluntary agreements, testing standards, and national security risk assessments. CAISI’s official website states it will establish voluntary agreements with private AI developers and conduct non-confidential assessments that may involve national security risks.

Meta states that its policy team is discussing the agreement details with the Department of Commerce and hopes to sign soon. The Department describes these interactions as routine work between CAISI and AI companies. The real disagreement is not whether to test a model once, but whether Meta’s future model releases will be incorporated into a more fixed government evaluation process after signing the agreement.

Meta Becomes the Most Notable Spectator

Meta has been brought to the forefront primarily because major U.S. AI labs are gradually establishing review collaborations with the government. According to media sources citing insiders, OpenAI, Anthropic, Google, xAI, and Microsoft have agreed to submit models to CAISI, while Meta has not yet signed a similar agreement. Since the full list has not been publicly disclosed, this should be viewed as media reports.

Frontier models are no longer just chatbots or content generation tools. The most capable systems may help researchers find software vulnerabilities, automatically generate code, and invoke external tools, but they could also be maliciously repurposed as attack aids. The government’s goal is to evaluate these capabilities’ boundaries before broader release.

Meta’s recent model plans have also attracted external attention. Reports indicate that in April, the company released a new model called Muse Spark, focusing on multimodal reasoning and tool use. Since details have not yet been fully verified in this round of public checks, a more cautious view is that Meta is still advancing its next-generation model capabilities, while the government hopes to include such models in pre-release evaluation frameworks.

For Meta, signing the agreement could reduce friction with the government but also introduce new uncertainties in release timing. The company needs to negotiate space regarding the scope of review, protection of trade secrets, access permissions, and release schedules.

A 30-Day Evaluation Window Is Not Approval, But It Will Affect Release Pace

The EO 14409 issued by the White House on June 2 provides a clearer institutional arrangement. The executive order requires relevant agencies to design a voluntary framework within 60 days, allowing developers to provide the federal government with access to “covered frontier models” for up to 30 days for safety assessments before broader release.

This executive order also states that it should not be interpreted as a mandatory license, pre-approval, or approval system. In other words, the U.S. government does not intend to turn frontier model releases into an official licensing regime; at least on paper, it emphasizes voluntary submission and safety evaluation.

However, the 30-day window is enough to alter the final phase of pre-release processes. For AI companies, model releases are increasingly like a compressed race lasting weeks or even days, where delays could impact market attention, enterprise trials, developer ecosystems, and funding narratives. Even if it’s not approval, any additional communication, modifications, or access restrictions resulting from government assessments could make release schedules harder to manage.

What remains unclear is whether companies will be required to delay releases, restrict access to certain users, or modify model capabilities if risks are identified during evaluation. Whether different companies’ models will be treated uniformly depends on the further implementation of the voluntary framework before early August.

Anthropic Incident Makes “Voluntary” More Sensitive

The Anthropic incident serves as a real stress test for this mechanism. According to AP, the U.S. government once asked Anthropic to prevent foreign nationals from using models Fable 5 and Mythos 5, and Anthropic complied by shutting down all customer access.

A key concern is that Amazon-related reports indicated Mythos could identify software vulnerabilities during testing. However, AP clarified that this does not mean the model can exploit vulnerabilities to launch attacks. Anthropic emphasized that such capabilities are not unique to its models; similar systems could also serve defensive purposes.

Axios reports that Amazon’s findings and the vulnerability issues heightened government concerns. Later, in an interview with Axios, Trump stated that he no longer viewed Anthropic as a security threat, and relations had eased.

This case leaves a direct question for the AI industry: if review agreements are nominally voluntary, can companies refuse to restrict access if the government deems certain capabilities a national security risk? If they cannot refuse, then “voluntary” assessments may effectively become constraints before release.

Meta’s Negotiation Stalls at the Boundary of Government Intervention

Meta’s current dilemma is that both options carry costs. Not signing the agreement means it remains the most cautious observer among major U.S. AI firms, easily compared to regulators and competitors. Signing, however, entails accepting a still-developing government evaluation process, which could introduce more uncertainty into future model releases.

If the final framework mainly focuses on safety testing, vulnerability sharing, and non-confidential assessments, AI companies are likely to incorporate it into pre-release compliance procedures. But if government evaluations lead to suspensions, access restrictions, or demands for modifications, the pace of public model releases will become even more unpredictable.

The impact of these negotiations extends beyond Meta. OpenAI, Google, Anthropic, xAI, Microsoft, and Meta are all accelerating their model capability races, with the U.S. government trying to move safety assessments earlier—before models are publicly released. The unresolved issue remains the boundary between government safety judgments and companies’ autonomy over releases.

Click to learn about job openings at Rhythm BlockBeats

Join the Rhythm BlockBeats official community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Twitter Official Account: https://twitter.com/BlockBeatsAsia