TL;DR
The US government has introduced a classified benchmarking process for advanced AI models, with voluntary pre-release evaluations and new oversight roles. This marks a significant shift in AI regulation toward secrecy and centralized control.
On June 2, President Trump signed Executive Order 14409, establishing a classified benchmarking process for advanced AI models and a voluntary framework for government pre-release assessments. This move shifts US AI oversight into a secretive domain, with the NSA and Treasury playing central roles, and marks a notable departure from previous hands-off approaches.
The order mandates that by August 1, 2026, the Treasury, NSA, and CISA, in coordination with other agencies, will set up a classified cyber-capability benchmark to evaluate AI models’ offensive capabilities. The NSA will decide which models qualify as covered frontier models, a designation that will be kept secret. Alongside this, a voluntary pre-release access framework will allow the federal government to evaluate AI models up to 30 days before their public release, with assessments shared with developers as appropriate.
Additionally, the order establishes an AI cybersecurity clearinghouse under the Treasury to share vulnerability intelligence between AI developers and critical infrastructure operators. It also allocates funding and personnel to improve AI vulnerability detection tools and cybersecurity talent. Participation in the pre-release framework is technically optional, but the designation as a trusted partner could influence federal procurement decisions, effectively creating a de facto requirement.
Implications of Classified AI Cybersecurity Benchmarks
This development represents a major shift in US AI governance, moving from voluntary, public standards to secretive, classified benchmarks that could influence market access and national security. By keeping evaluation criteria secret, the US aims to prevent adversaries from gaming the system but risks reducing transparency and international cooperation. The move signals an increased focus on national security at the expense of open, contestable standards, contrasting with European approaches like the EU AI Act, which emphasizes transparency and public thresholds.
As an affiliate, we earn on qualifying purchases.
From Voluntary Frameworks to Secrecy in AI Regulation
Earlier efforts at AI regulation in the US favored voluntary standards and transparency, with some attempts to impose mandatory testing. The executive order marks a significant policy shift, partly driven by concerns over AI capabilities’ potential misuse and competition. The order is a second attempt after an earlier version was reportedly pulled due to fears it would hinder US competitiveness. It reflects a broader trend of centralizing oversight, with the NSA and Treasury gaining new roles in AI security, which previously had minimal involvement in AI governance.
AI model vulnerability detection software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Unresolved Questions About Implementation and Oversight
It remains unclear how the NSA will define and enforce the classified benchmarks, and what specific criteria will be used to designate a model as a covered frontier model. The scope of government access to proprietary data and the legal protections for developers participating in the pre-release assessments are also still under discussion. Additionally, the long-term impact of this secrecy on international cooperation and AI innovation is uncertain, as other nations may adopt different transparency standards.
AI pre-release evaluation platforms
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Next Steps in US AI Security Policy Development
Developers and industry stakeholders will need to decide whether to participate in the voluntary pre-release assessments by August 1, 2026. The NSA and Treasury will finalize the classified benchmarks and designation process, with the first evaluations likely to occur shortly after the deadline. Congressional and industry debates about the balance between security and transparency are expected to intensify, potentially influencing future legislation. Monitoring how the framework is implemented and its effects on AI development will be critical in the coming months.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is the main purpose of the classified AI benchmarks?
The benchmarks aim to evaluate the offensive cyber capabilities of advanced AI models while keeping the criteria secret to prevent adversaries from gaming the system or developing countermeasures.
Will participation in the pre-release assessments be mandatory?
Participation is technically voluntary, but the designation as a trusted partner and potential preference in federal procurement could effectively make it a requirement for vendors seeking government contracts.
How does this differ from European AI regulations?
The US approach involves classified, secret benchmarks, whereas the European Union emphasizes public, contestable thresholds like compute limits, promoting transparency and international cooperation.
What are the risks of keeping benchmarks classified?
Classified benchmarks may reduce transparency, hinder independent verification, and potentially allow biases or inaccuracies to go unchallenged, affecting accountability and trust in AI safety standards.
What happens if a developer refuses to participate?
Refusing to participate may limit access to federal contracts and trusted partner status, possibly affecting market opportunities within government procurement channels.
Source: ThorstenMeyerAI.com
