According to media reports, the Trump administration is urging Meta to submit cutting-edge AI models for voluntary review by the federal government, with Meta being one of the few major U.S. AI developers that has not yet signed a similar arrangement. For AI companies, this is not a traditional approval process, but it will affect how the most powerful models undergo security testing before public release, what capabilities the government can see in advance, and whether model access may be required to be adjusted when it involves national security risks such as cyber, bio, or chemical weapons. Companies such as OpenAI, Anthropic, Google, xAI, and Microsoft are said to have agreed to collaborate with a Commerce Department agency, which has magnified Meta's negotiation progress.

This engagement primarily revolves around the Center for American Innovation and Security in AI (CAISI). CAISI is part of the Commerce Department and NIST system, reorganized in June 2025 from the U.S. AI Safety Institute during the Biden era, with a stronger focus on voluntary agreements, testing standards, and national security risk assessments. The CAISI website states that it will establish voluntary agreements with private AI developers and conduct non-classified assessments that may involve national security risks.

Meta has stated that its policy team is in discussions with the Commerce Department on the details of the agreement and hopes to sign soon. The Commerce Department describes the engagement as part of CAISI's regular work with AI companies. The real disagreement is not over testing a one-off model but whether, after signing the agreement, Meta's future model releases will be incorporated into a more formalized government evaluation process.

Meta Becomes the Most Prominent Observer

Meta has been thrust into the spotlight, primarily because major U.S. AI labs have been progressively establishing review collaborations with the government. According to sources, OpenAI, Anthropic, Google, xAI, and Microsoft have agreed to submit models to CAISI, while Meta has not yet signed a similar agreement. As the official full list has not been disclosed, this statement should still be considered media speculation.

Cutting-edge models are no longer just chatbots or content generation tools. The most powerful systems could help researchers uncover software vulnerabilities, auto-generate code, invoke external tools, or be weaponized by malicious actors. The government is keen on preemptively assessing these boundary capabilities before models are more widely disseminated.

Meta's recent model plans have also attracted external attention. Reportedly, the company released a new model called Muse Spark in April, focusing on multimodal reasoning and tool usability. As specific details have not been fully confirmed in this round of public scrutiny, a more cautious statement would be that Meta is still advancing the capabilities of next-generation models, and the government hopes to incorporate such models into a pre-release evaluation framework.

For Meta, signing the agreement helps reduce friction with the government and may also introduce new uncertainties regarding releases. The company needs to negotiate aspects such as the scope of review, protection of business secrets, access rights, and release timing.

A 30-Day Assessment Window Is Not Approval, but Will Affect Release Pace

Executive Order 14409 issued by the White House on June 2 provides a more specific institutional arrangement. The executive order requires relevant agencies to design a voluntary framework within 60 days, allowing developers to provide the federal government with up to 30 days of access to "covered frontier models" for security assessment before broader release.

This executive order explicitly states that it should not be interpreted as a mandatory licensing, pre-approval, or approval system. In other words, the U.S. government has not transformed the release of frontier models into a formal licensing regime, at least emphasizing voluntary submission and security assessment at the textual level.

However, the 30-day window is sufficient to alter the final stage of the pre-release process. For AI companies, model release is increasingly resembling a competition compressed into weeks or even days, where delays can affect market momentum, enterprise customer trials, developer ecosystems, and funding narratives. Even if not an approval process, as long as government evaluation leads to additional communication, modifications, or access restrictions, release schedules become more challenging to manage.

It is currently unclear whether, if risks are identified during the evaluation, companies must postpone releases, restrict access for some users, or adjust model capabilities. Whether models from different companies will be handled according to a unified standard also awaits further development of the voluntary framework by early August.

The Anthropic Incident Makes "Voluntary" More Sensitive

The Anthropic incident serves as the most realistic stress test for this mechanism. AP reported that the U.S. government had requested Anthropic to prevent foreign nationals from using the Fable 5 and Mythos 5 models, leading Anthropic to shut down all customer access for compliance.

One focal point of concern was an Amazon report indicating that Mythos could identify software vulnerabilities during testing. However, AP also noted that this did not mean the model was capable of exploiting vulnerabilities to launch attacks. Anthropic emphasized that such capabilities are not unique to its models alone, and similar systems could also serve defensive purposes.

Axios reported that Amazon's report and vulnerability issues have heightened government concerns. Trump later stated in an Axios interview that Anthropic is no longer seen as a security threat, and relations between the two parties have since improved.

The case has raised a very direct question for the AI industry: If a review agreement is ostensibly voluntary, can a company refuse to restrict access once the government deems a certain capability a national security risk? If refusal is not an option, the so-called voluntary assessment would effectively resemble pre-release constraints in operation.

Meta Negotiations Stalled at Government Intervention Boundary

Meta is currently in a difficult position where both choices come with a cost. Without signing the agreement, it will continue to be the most prominent observer among the major AI companies in the U.S., making it an easy target for regulators and competitors to compare. By signing the agreement, the company will have to accept a government evaluation process that is not yet fully developed, potentially adding another layer of uncertainty to future model releases.

If the final framework primarily focuses on security testing, vulnerability sharing, and non-confidential assessments, AI companies are likely to incorporate it into their pre-release compliance process. If the government evaluation can lead to release suspensions, access restrictions, or remediation requirements, the pace of releasing cutting-edge models will be even more unpredictable.

The impact of this negotiation extends beyond just Meta. OpenAI, Google, Anthropic, xAI, Microsoft, and Meta are all accelerating the competition in model capabilities, while the U.S. government is attempting to advance the security evaluation of the most powerful models before public release. What remains truly unclear is the boundary between government security assessments and a company's autonomy to release.

Welcome to join the official BlockBeats community:

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

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

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