White House keeps new AI evaluation framework secret

The tension between national security classification and technical standardization is a recurring pattern in dual-use technology governance. Historically, export control regimes and intelligence community evaluations have operated under similar secrecy constraints to prevent adversarial learning, often resulting in fragmented compliance landscapes where only cleared entities understand the rules. In AI governance specifically, this mirrors earlier debates around cryptographic standards and vulnerability disclosure processes, where government-held criteria created information asymmetries between regulators and the research community. Voluntary safety frameworks in emerging technologies frequently struggle with this trade-off; making criteria public invites gaming and adversarial adaptation, while keeping them private undermines the legitimacy and interoperability required for broad adoption. This dynamic is particularly acute when domestic frameworks interact with international regulatory regimes like the EU AI Act, creating potential misalignments between transparent statutory requirements and opaque executive evaluations. The current situation reflects this structural friction between operational security needs and the norms of open technical standard-setting.

On August 4, 2026, Axios reported that the White House intends to keep its newly completed framework for evaluating advanced artificial intelligence models secret from the public, according to three sources familiar with the discussions. The administration conducted staff-level meetings with select industry representatives two days prior to review the finalized criteria, but companies not invited to these briefings remain unaware of the framework's specific contents or requirements. According to the reporting, this voluntary framework carries global implications for AI security, yet its details will be restricted exclusively to participating companies and designated trusted partners.

The decision to withhold the framework has generated significant concern among observers regarding transparency and accountability. Critics argue that keeping the evaluation criteria private leaves policymakers, researchers, and U.S. allies outside the process unable to understand how the administration plans to implement a key pillar of its AI policy. Axios reports that this secrecy prevents external stakeholders from preparing for new standards or holding the evaluation process accountable. Conversely, the administration maintains that restricting access is necessary, though specific justifications cited in the available sources focus on national security risks and the need for effective coordination within high-security environments during model reviews.

Substantively, the framework defines a "covered frontier model" as a closed-source system possessing state-of-the-art capabilities and presenting national security risks, according to multiple sources briefed on the White House meetings. However, Axios notes that there is currently no clear public definition of what constitutes "state-of-the-art" capabilities or a specific "national security risk" under this regime. Crucially, the framework explicitly excludes open models from its testing and evaluation requirements. Sources indicate the document states that nothing within it should be interpreted as restricting open models once they have been released, effectively exempting them from the pre-release government review process applicable to closed frontier systems.

For models that do fall under the framework's purview, the evaluation process involves strict security protocols. During a mandated 30-day pre-release government review period, employee access to the models would be limited, and the systems would be stored in high-security environments with detailed logging of all access. The review itself will involve various administration officials rather than being centralized within a single office or agency. This distributed approach suggests an interagency effort to assess compliance and risk, though the lack of public documentation makes it difficult for external parties to verify the rigor or consistency of these assessments.

International reaction to the secretive nature of the framework remains uncertain due to the lack of disclosed criteria for partnership. It is currently unclear which entities qualify as "trusted partners" eligible for early access to advanced models under the new rules. When approached for comment regarding whether it qualifies as a trusted partner under the undisclosed framework, the European Union declined to comment, and the United Kingdom did not respond to multiple requests. This diplomatic silence highlights the ambiguity facing even close allies who may be subject to or beneficiaries of the framework without knowing the governing terms.

The controversy arises against a backdrop of heightened cybersecurity concerns and rapid technological advancement by Chinese AI developers, which has reignited debates over how to manage different classes of AI models. By distinguishing between closed frontier models and open models, the White House appears to be adopting a bifurcated regulatory strategy that subjects proprietary high-capability systems to government scrutiny while leaving open-weight ecosystems largely untouched by this specific evaluation mechanism. However, because the definitions driving this distinction remain classified or internal, the practical boundary between regulated and unregulated development remains opaque to the broader AI community.

Companies excluded from private briefings cannot prepare for compliance, risking delayed product releases during the 30-day review window. International allies face uncertainty about trusted partner status, potentially fragmenting transatlantic AI governance. Open model developers gain temporary exemption but face long-term regulatory ambiguity if definitions of 'frontier' expand. The secrecy undermines trust in voluntary regimes, potentially pushing stakeholders toward binding legislation or alternative standards bodies. Research communities lose visibility into government safety benchmarks, hindering independent validation of AI risk assessments.