President Trump signed a national security presidential memorandum on Friday ordering the US military and intelligence agencies to accelerate their adoption of cutting-edge AI. The directive, NSPM-11, establishes a framework for “rapid onboarding of the most advanced AI models from multiple vendors.” It also bars any company from disabling, degrading, or modifying an AI system that warfighters depend on without prior government approval.

That vendor restriction is the most striking provision. It means an AI company cannot pull a deployed model from military use unilaterally, even if the company has safety concerns about how it is being used. The clause lands directly in the context of the Pentagon’s ongoing feud with Anthropic, which was blacklisted as a supply chain risk after refusing to allow its Claude models to be used for autonomous weapons or mass surveillance.

“The men and women who defend our nation deserve the best, most secure and most reliable AI in the world,” said Michael Kratsios, director of the White House Office of Science and Technology Policy.

The memo also directs Defence Secretary Pete Hegseth to issue an updated directive on autonomous weapon systems within 90 days. That update would revise DoD Directive 3000.09, the foundational Pentagon policy governing when and how autonomous and semi-autonomous weapons can be used, including requirements for human judgment before lethal force is applied.

There are stated limits. The memo prohibits defence agencies from creating or releasing AI models designed to “censor free speech, embed ideological bias or conduct unlawful surveillance against the American people.” But it does not define those terms or explain how compliance would be enforced.

The directive follows Tuesday’s executive order that established a voluntary 30-day review window for frontier AI models before public release. Together, the two documents outline a dual approach: light-touch regulation for the commercial sector and aggressive adoption for the military.

The “multiple vendors” language signals a shift away from single-provider dependency. Until recently, Anthropic was the only AI vendor approved for classified military use. After the Pentagon signed classified deals with Nvidia, Microsoft, and AWS, the administration is now formalising a multi-vendor approach.

The memo makes accountability central. Commanders, directors, and agency heads remain responsible for ensuring AI is used in line with its stated obligations. Annual reviews of key guidance across the national security enterprise are required to keep pace with the AI frontier. Whether those reviews will be meaningful or performative remains an open question.

President Trump signed an executive order on Tuesday establishing a voluntary framework for government review of frontier AI models before public release, ending weeks of internal White House conflict over how aggressively to regulate the technology. The order, titled “Promoting Advanced Artificial Intelligence Innovation and Security,” was signed privately without the usual livestream or public ceremony, a contrast with the fanfare that typically accompanies presidential AI announcements.

The final version is substantially narrower than the draft Trump rejected on 21 May, when he scrapped a planned signing ceremony over concerns that the order “could dull America’s edge on AI technology.” The original draft proposed a 90-day mandatory pre-release review period and would have given the government formal evaluation authority over frontier models. The signed version asks companies to voluntarily submit models 30 days before release and participate in a collaborative framework rather than submitting to mandatory testing.

What the order does

The executive order establishes three main mechanisms. First, a voluntary pre-release review framework in which AI developers can engage the government to determine whether models under development qualify as “covered frontier models,” provide access for up to 30 days before planned release, and collaborate on selecting “trusted partners” for early access. The framework is explicitly voluntary, meaning companies can decline to participate without penalty.

Second, the order creates an AI cybersecurity clearinghouse within 30 days, coordinated by the Treasury Secretary, the National Cyber Director, the NSA, and CISA. The clearinghouse will scan for software vulnerabilities, validate discoveries, and coordinate remediation and patch distribution, a direct response to the Mythos crisis that demonstrated how AI-discovered vulnerabilities can outpace existing disclosure and patching processes.

Third, federal agencies are directed to develop benchmarks for assessing AI models’ cybersecurity capabilities and to strengthen the government’s own security defences against AI-enabled threats. The order also addresses AI safety research, though the specific provisions are less prescriptive than what the original draft contained.

What was cut

The differences between the scrapped draft and the signed order reflect the victory of the pro-industry faction within the White House. The 90-day mandatory review was reduced to a 30-day voluntary window. The formal government evaluation authority was replaced with a collaborative framework. The reporting requirements for companies developing powerful models, which would have echoed provisions in Biden’s repealed AI executive order, were softened to avoid what industry allies characterised as regulatory overreach.

Silicon Valley’s objections to the original draft were decisive. AI companies argued that mandatory pre-release testing would slow American innovation, create a competitive disadvantage relative to Chinese firms facing no equivalent requirements, and establish a precedent for government gatekeeping of technology deployment. The signed order addresses those concerns by making participation voluntary and framing the government’s role as collaborative rather than regulatory.

The gap it leaves

The voluntary framework means the order’s effectiveness depends entirely on whether AI companies choose to participate. Companies already engaged in pre-release testing with CAISI, including Google, Microsoft, and xAI, may continue or expand that cooperation. Companies that view government review as commercially disadvantageous or that are racing to ship products can simply opt out.

The EU’s AI Act, entering full enforcement in August, provides a stark contrast: mandatory requirements, statutory authority, and penalties for non-compliance. The Trump order establishes norms and creates institutional infrastructure (the cybersecurity clearinghouse, the benchmark development process) but relies on goodwill rather than obligation.

For the White House, the quiet signing may be the point. The order gives the administration a policy document it can reference when asked about AI oversight, creates structures that could be strengthened later, and avoids a public confrontation with an AI industry whose leaders are among the administration’s most visible supporters. Whether a voluntary framework is adequate for a technology that can discover 10,000 zero-day vulnerabilities in a month is the question the order deliberately leaves unanswered.

President Donald Trump signed an executive order on Tuesday designed to give the government a chance to review powerful AI models before they are released. 

The order asks certain AI companies to voluntarily submit their new models to the government for testing or evaluation 30 days before releasing the products to the public. A previous draft of the order had called for a voluntary review up to 90 days in advance, though AI industry insiders had pushed for something closer to a two-week window.

Trump had been slated to sign the more demanding version of the order in late May, but delayed after industry pushback, including from venture capitalist and former White House AI czar David Sacks. The president said at the time that he didn’t want to do anything to get in AI firms’ way of leading against China. 

“Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models,” reads the order, published Tuesday. 

Trump had planned to sign the EO with a bevy of Silicon Valley’s top CEOs in attendance, but ended up signing the current version privately.

In addition to the voluntary governmental AI model review, the EO directs the Department of Justice to treat crimes like AI-assisted hacking and unauthorized access as a high-priority enforcement area. 

This isn’t the president’s first EO on AI. Last December, Trump signed an order directing the development of “one rulebook,” or a national AI policy framework, intended to preempt state AI laws.