Security

US Executive Order Sets Up Voluntary, Classified Cyber Testing of Frontier AI Models Before Release

An order signed on 2 June asks US security agencies to build a classified benchmark for AI models' cyber capabilities and lets developers give the government up to 30 days' access before release. Participation is voluntary.

Abstract illustration of a shield over a hexagonal lattice, in deep blue and cyan.
Abstract illustration of a shield over a hexagonal lattice, in deep blue and cyan.

President Donald Trump signed an executive order on 2 June 2026 that creates a voluntary route for AI developers to have their most capable models assessed by US national security agencies for advanced cyber capabilities before those models are released. The order, titled “Promoting Advanced Artificial Intelligence Innovation and Security”, was published in the Federal Register on 5 June as Executive Order 14409.

Alongside the testing framework, the order sets a series of 30-day deadlines for strengthening the defence of federal systems and creates a clearinghouse for finding and fixing software vulnerabilities with help from the AI industry.

The testing framework

The core of the order is Section 3, headed “Secure Frontier Model Deployment”. Within 60 days, the Treasury, the Department of War (through the National Security Agency) and the Department of Homeland Security (through the Cybersecurity and Infrastructure Security Agency, CISA) must develop a classified process to assess what the order calls “the advanced cyber capabilities of AI models”. They are to consult the National Cyber Director, the President’s science and technology adviser and the Commerce Department through NIST.

That process also sets the threshold for a new category, the “covered frontier model”. The order gives the decision on whether a model meets it to the Director of the NSA.

Developers can then opt into a voluntary framework under which they may:

  • ask the government whether a model they are building meets the “covered frontier model” threshold;
  • give the government access to a covered model for up to 30 days before releasing it to other trusted partners, subject to confidentiality, cybersecurity, insider-risk and intellectual-property protections; and
  • work with the government to select “trusted partners” who receive early access to covered models “to promote secure innovation and strengthen the cybersecurity of critical infrastructure”.

The order is explicit that this is not a permission system. “Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement,” it states, covering the development, release and distribution of AI models, including frontier models.

Hardening federal systems

Section 2 directs a set of near-term actions, most of them due within 30 days:

Who What Deadline
Committee on National Security Systems Prioritise the cyber defence of National Security Systems 30 days
Secretary of War Prioritise the cyber defence of Department of War systems 30 days
DHS, through CISA Issue Binding Operational Directives and guidance; expand AI-enabled defensive tools; widen access to cyber tools, including covered frontier models where appropriate 30 days
Treasury Form a voluntary AI cybersecurity clearinghouse with the AI industry and critical infrastructure operators 30 days
Office of Management and Budget Identify grant funding that could support advanced AI vulnerability detection 30 days
Office of Personnel Management Expand cybersecurity hiring and placement pathways 60 days

The clearinghouse is the most concrete new body. According to the order, it will coordinate and deconflict scanning for software vulnerabilities, discover and validate them, and prioritise the distribution of patches. CISA’s guidance is meant to extend access to cyber tools beyond federal agencies to state and local authorities and to critical infrastructure operators, which the order illustrates with rural hospitals, community banks and local utilities.

Section 4 tells the Attorney General to prioritise enforcement of existing federal computer-crime, identity-fraud and wire-fraud statutes against anyone who uses AI to access or damage computers without authorisation, or to further other crimes.

How the White House frames it

An accompanying White House fact sheet presents the order as a partnership with industry rather than regulation. “As AI capabilities increase, new cyber and national security considerations emerge that require coordinated action,” it says, adding that “government should partner with innovators – not stifle them.” The fact sheet does not name any AI companies or models.

What it means

Analysis. Three features of the order stand out.

It is about cyber capability, not AI safety in general. The benchmark measures one thing: how good a model is at offensive and defensive cyber work. Other categories often discussed in frontier AI testing, such as biological or chemical misuse, are not mentioned. That narrows the scope, but it also targets the capability that most directly affects the security of networks the government is responsible for.

The criteria will not be public. Because the benchmarking process is classified and the designation rests with the NSA, developers and outside observers will not necessarily know where the line for a “covered frontier model” sits. Whether the government shares assessments is left to discretion: the order says they may be shared with developers and researchers “as appropriate”.

The 30-day window is early access, not approval. The government’s access period comes before release to other trusted partners, and the order rules out any preclearance requirement. In practice the framework looks designed to put the most capable models in front of defenders, including federal agencies and critical infrastructure operators, before they are widely available, rather than to gate releases.

Because participation is voluntary, its effect will depend on which developers take part and on how much each side is willing to share.

What to watch

  • Early July: the 30-day deadlines for CISA’s directives, the Treasury clearinghouse and the OMB grant review.
  • Around 1 August: the 60-day deadline for the classified benchmark and the voluntary framework, and whether any part of either is made public.
  • Participation: which AI developers say they will engage, and how “trusted partners” are chosen.

Sources

See All

Type to search. ↑ ↓ to navigate, ↵ to open.