Trump’s frontier AI plan leaves public safeguards unanswered

President Donald Trump’s June 2 executive order on AI is best understood not as comprehensive AI regulation, but as a national security and cybersecurity intervention focused on advanced models with potentially consequential cyber capabilities.

It does not create a licensing, preclearance, or permitting system; impose universal testing obligations; or establish enforceable new rights for people harmed by AI. Nor does it create broader privacy, civil liberties, consumer protection, or anti-discrimination safeguards.

What is does do is direct federal agencies to harden federal systems against AI-enabled threats, create a clearinghouse for software vulnerabilities, and establishe a voluntary process through which companies can give the government early access to powerful models with advanced cyber capabilities.

That architecture moves federal AI policy closer to the development pipeline of frontier systems while preserving the administration’s core political commitment to avoiding rules it considers burdensome, a not-so-subtle cue to lawmakers regarding legislation that would ban states from regulating how AI is used and deployed.

The White House says the U.S. must deploy advanced technology rapidly, protect American intellectual property, and strengthen defenses against external threats.

The order’s operative language makes national security, not consumer protection, civil rights, or general AI accountability, the basis for government engagement with the private model developers at the center of the industry.

The final version of the order reflects a compromise reached after internal administration debate. Trump had delayed a planned May 21 signing, saying he did not want to do anything that could impede the U.S.’ lead over China.

The order instead asks companies to work with national security agencies while expressly prohibiting agencies from treating the framework as a licensing, preclearance, or permitting system.

The immediate operational focus is cybersecurity. Within 30 days, the Committee on National Security Systems (CNSS) and the secretary of defense are directed to prioritize cyber defense of national security and Defense Department information systems.

CNSS is a U.S. federal interagency body that establishes baseline cybersecurity policies, directives, and standards governing the protection of national security systems across the U.S. government

The same deadline directed the secretary of the Department of Homeland Security, through the Cybersecurity and Infrastructure Security Agency (CISA) and in consultation with the Office of Management and Budget (OMB), the national security adviser, and the national cyber director, to issue binding operational directives and guidance.

These directives aim to strengthen civilian federal cybersecurity, expand programs using AI-enabled defensive tools, and provide federal, state, local, and critical infrastructure operators with access to cyber tools, potentially including covered frontier models.

The order treats AI as a defensive capability that could be extended beyond the federal enterprise to organizations whose compromise could have outsized local consequences. Whether those organizations can realistically absorb, staff, and securely operate these tools is a separate question the order does not answer.

Separately, CISA’s June 10 Binding Operational Directive 26-04 (BOD-26-04) is the clearest enforceable civilian agency measure issued following the executive order’s call for binding directives and guidance. CISA itself linked the directive to AI-enabled acceleration of vulnerability discovery and exploitation.

Rather than relying principally on static severity scores, it requires agencies to prioritize remediation using four risk factors: whether a vulnerable asset is exposed, whether the flaw is known to have been exploited, whether exploitation can be automated, and the degree of technical control an attacker could obtain after compromise.

The directive is distinct from the executive order’s frontier model review framework. BOD 26-04 governs how civilian agencies identify and remediate vulnerabilities already present in their systems, while the separate frontier model program is intended to assess whether advanced AI models could materially accelerate cyber operations before broader release.

Together, they reflect the administration’s concern that AI may shorten the interval between finding a vulnerability, building an exploit, and gaining operational access to a target.

The executive order also gives the Treasury Department an unusual cyber role. In consultation with the national cyber director, National Security Agency (NSA), and CISA, Treasury is directed to form an AI cybersecurity clearinghouse with voluntary participation from AI companies and critical infrastructure operators.

The clearinghouse is supposed to coordinate and deconflict vulnerability scanning, discovery, and validation, then help prioritize remediation and patch distribution. It responds to a real weakness in the American cyber ecosystem.

Vulnerability intelligence, exploit reporting, and patching are fragmented across vendors, agencies, and sectors, even as capable AI systems can help defenders analyze large volumes of security data and help attackers identify weaknesses at greater speed.

Still, a clearinghouse is not a self-executing solution. Its effectiveness will depend on what information firms share, who can see it, whether participants can exchange sensitive technical data without creating a new target for espionage or intrusion, and how conflicts are resolved when a company wants to delay disclosure or remediation.

The order says collaboration is voluntary and provides no public governance structure, reporting mechanism, or independent oversight process for the clearinghouse.

OMB was also directed only to determine whether relevant existing grant money could be directed toward advanced AI vulnerability detection, while the Office of Personnel Management was ordered to expand a federal cybersecurity hiring pathway.

Neither provision guarantees new appropriations or a workforce large enough to meet the mission.

The order’s cyber provisions lead directly to its more consequential mechanism: a classified process for identifying which AI models warrant special national security attention.

The center of gravity lies in Section 3, which orders Treasury, NSA, and CISA, in consultation with the national cyber director, White House technology officials, National Institute of Standards and Technology, and others, to create and maintain a classified benchmarking process.

The purpose is to evaluate advanced cyber capabilities in AI models and set the threshold for the new “covered frontier model” designation. Such a determination will be made by the NSA director after consultation. It does not publicly spell out a model-by-model designation or appeal process.

Agencies may share assessments with developers and researchers as appropriate, but the underlying benchmarking process will be classified, a move that gives the government latitude to examine sensitive capabilities without publicly disclosing how it evaluates models that could assist vulnerability discovery, exploitation, code generation, or automated intrusion.

It also creates an accountability problem. The public, independent researchers, and even smaller AI developers will have limited ability to understand where the threshold lies, whether it is being applied consistently, or whether it is responsive to threats other than the ones intelligence agencies prioritize.

Classification may be justified for some details, but the order does not require public criteria, an unclassified summary of outcomes, outside red team review, or a mechanism for challenging a designation.

Once a model is identified as potentially consequential, the order turns from assessment to access: how much visibility the government can obtain before release and who else may be included.

The framework that follows is nominally voluntary. Developers can ask the government whether a model under development qualifies as a covered frontier model. They can provide access to such a model for up to 30 days before releasing it to other trusted partners, subject to confidentiality, cybersecurity, insider risk, and intellectual property protections.

They can also work with the federal government to select trusted partners that receive early access to promote secure innovation and strengthen critical infrastructure.

The order makes clear the White House wants early visibility into a narrow category of high-capability systems without assuming legal authority to stop their release. It is also an effort to turn informal information sharing relationships into a more regular national security practice.

Reuters reported that OpenAI and Anthropic already had been working with the government to test unreleased models, while Google, Microsoft, and xAI had agreed in May to provide early access for national security evaluations.

Reuters also reported that Meta was the only major U.S. developer without such an agreement.

The executive order contemplates government involvement in selecting “trusted partners” for early access but says little about how trust will be defined or how excluded researchers, competitors, public interest groups, or smaller infrastructure operators could be treated.

Early access can strengthen defenses by allowing vetted organizations to prepare for risks before a model is widely distributed. It can also create a privileged circle around the most consequential technology, deepen the advantage of established firms, and present difficult questions about equitable access, trade secrets, and security clearance-like standards without making those standards visible.

The order contains no binding privacy, civil rights, consumer protection, labor, or anti-discrimination requirements. It does not require developers to publish safety evaluations, document training data practices, explain automated decisions, or report model failures outside the cyber context.

Section 4 of the order tells the attorney general to prioritize enforcement of existing federal laws against people who use AI to gain unauthorized access to or damage computers, steal data, or further crimes.

The order specifically mentions AI agents used to unlawfully access data but does not change the elements of the Computer Fraud and Abuse Act, wire fraud statutes, or identity-related criminal laws.

Three days after the executive order, the White House issued National Security Presidential Memorandum-11 (NSPM-11) on AI in the national security enterprise.

The memorandum is broader and more direct about military and intelligence use. It calls for rapid adoption of commercial and open source AI, accelerated procurement, and systems that are reliable, robust, steerable, and controllable.

It also says national security AI must not be used to censor speech, embed ideological bias, or conduct unauthorized or unlawful surveillance, and must remain consistent with constitutional protections and privacy rules.

The memorandum further directs relevant agencies, to the maximum extent allowed by law, to terminate contracts with companies that repeatedly act contrary to its AI-use policies. That is a significant enforcement lever and complicates any impression that the memorandum is only aspirational internal guidance.

Like the executive order, NSPM-11 also states that it creates no substantive or procedural right that can be enforced against the government. The policy therefore asks the agencies expanding AI use to police themselves through contracting, guidance, training, and internal accountability.

While that may be essential for speed and flexibility, it leaves Congress, inspectors general, courts, and the public with fewer formal tools for testing whether the promised protections are being honored.

Cybersecurity expert Chuck Brooks, writing in GovCon Wire, argues that cybersecurity has become the foundation of AI competitiveness and highlights the need for security across model development, identity, data, supply chains, and autonomous AI agents.

The security problem does not end with whether a model can generate exploit code. It includes whether model weights can be stolen, training data can be poisoned, agent permissions can be abused, and critical systems can remain controllable under attack.

The order recognizes parts of that challenge, particularly vulnerability discovery and critical infrastructure defense, but it leaves much of the security architecture as an implementation problem for agencies and companies.

The first tests will come quickly. The order’s 30-day deadlines fell on July 2, and the classified benchmarking process and voluntary frontier model framework are due by August 1.

The public likely will never see the benchmark, but should be able to judge whether the administration explains who participates in the voluntary process, what protections govern government access to proprietary models, how trusted partners are selected, and how cybersecurity benefits reach the smaller critical infrastructure organizations the order identifies.

Those answers will determine whether the order becomes a disciplined national security channel into frontier AI or an opaque partnership that gives the federal government greater access to powerful models while leaving the public little visibility into the rules that govern them.

Article Topics

cybersecurity  |  legislation  |  national security  |  U.S. AI policy  |  U.S. Government