AI Benchmarks As National Security Instruments: The Washington Strategy Unveiled

On June 2, President Biden signed Executive Order 14409, establishing a classified benchmarking process to evaluate the cyber capabilities of advanced AI models. This process will determine which models are designated as “covered frontier models,” a label with significant national security implications, and will be managed by the NSA, Treasury, and other agencies. The order also introduces a voluntary framework allowing developers to share their models with the federal government for pre-release evaluation, marking a shift toward increased government oversight of AI systems.

The executive order mandates the creation of a classified cyber-capability benchmark for AI models, with the NSA Director responsible for designating models as “covered frontier models” by August 1, 2026. Alongside this, a voluntary pre-release access program will allow developers to share their models with federal agencies for up to 30 days before public deployment, with evaluations shared as appropriate. The order also establishes an AI cybersecurity clearinghouse under the Treasury Department to facilitate information sharing on vulnerabilities between the AI industry and critical infrastructure operators. Additionally, funding and staffing are allocated to enhance AI vulnerability detection and federal cybersecurity talent.

Legal analysts note that participation in the pre-release program is opt-in, but the designation as a trusted partner—linked to participation—may become a key factor in federal procurement decisions. The order emphasizes voluntary cooperation, but past actions, such as the suspension of an AI model by the government over cyber capabilities, suggest that the government may enforce compliance when necessary. The benchmarks will be classified, raising concerns about transparency and the potential for opaque decision-making, contrasting with Europe’s public and contestable risk thresholds.

Implications of Classified AI Cybersecurity Benchmarks

This development signifies a major shift in US AI governance, moving away from voluntary, open standards toward classified, government-controlled benchmarks. While intended to enhance national security by assessing AI cyber capabilities, it raises concerns about transparency, accountability, and the potential for opaque decision-making. The move also signals increased government interest in AI as a strategic asset, with the possibility that trusted partnership status could influence federal procurement and market access for AI vendors. This approach contrasts sharply with European efforts, which favor transparent, public risk thresholds, highlighting a divergence in global AI regulation strategies.

Background on US AI Regulation and Security Measures

Earlier in 2026, the US government took steps to regulate AI through executive actions, including a move to suspend access to advanced AI models exhibiting significant cyber capabilities. This marked a departure from previous hands-off policies, reflecting growing concerns over AI’s national security implications. The current executive order builds on this foundation, formalizing a classified benchmarking process that evaluates AI models’ cyber capabilities, aligning with broader efforts to integrate AI into national security frameworks. The order is a second attempt after an earlier version was reportedly withdrawn due to fears it would hinder US competitiveness.

Internationally, the European Union has adopted a different approach with its AI Act, setting public risk thresholds based on training compute, which are openly contestable and transparent. The US’s move toward classified benchmarks represents a strategic divergence, emphasizing secrecy over openness in AI governance.

Uncertainties Surrounding the Classified Benchmark System

It remains unclear how the classified benchmarks will be developed, what specific cyber capabilities they will measure, and how the NSA will enforce designations. The criteria will be secret, raising concerns about potential biases, inaccuracies, or manipulation. Additionally, the extent to which participation in the pre-release framework will influence market access or government contracts is still uncertain. The long-term impact on AI innovation and international competitiveness also remains to be seen, especially given the contrasting European approach.

Next Steps for US AI Security and Industry Engagement

Developers and AI firms will need to decide whether to participate in the voluntary pre-release framework before the August 1 deadline. The government is expected to finalize the classified benchmarks and begin designating models shortly after. Congressional debates may follow on whether to make such assessments mandatory or to increase transparency. Meanwhile, industry and security experts will monitor how the classified benchmarks influence AI development, market access, and international regulatory alignment. The Biden administration may also expand or refine the framework in response to emerging threats or industry feedback.

Key Questions

What is the purpose of the classified AI benchmarks?

The benchmarks are intended to evaluate the cyber capabilities of advanced AI models to determine their national security risk and designate certain models as “covered frontier models” for oversight.

Will companies be required to participate in the pre-release access program?

No, participation is voluntary, but being designated as a trusted partner through participation could influence federal procurement and market access.

How does this US approach compare to Europe’s AI regulation?

The US is implementing classified, opaque benchmarks, while Europe favors transparent, public risk thresholds such as compute-based limits.

What are the potential risks of classified benchmarks?

Classified benchmarks could lack transparency, be subject to bias or inaccuracies, and limit external review or challenge, raising concerns about accountability.

What happens if a developer refuses to share their model for evaluation?

The order does not specify mandatory sharing, but refusal may affect their eligibility for trusted partner status and federal contracts.

Source: ThorstenMeyerAI.com