When President Trump pulled the planned signing of an AI cybersecurity executive order on May 22, his explanation was four words long: "I postponed it." The draft itself, obtained and published by Politico over the weekend, is rather more detailed. Across seven pages it lays out the most concrete federal scheme yet floated for reviewing advanced AI models before they reach the market, and its single most consequential sentence is the one explaining what the scheme is not.
The order, titled "Promoting Advanced Artificial Intelligence Innovation and Security," would have set three policy tracks in motion. The Cybersecurity and Infrastructure Security Agency would have 30 days to issue binding directives extending AI-enabled defensive tools to federal systems and to a specific list of underserved infrastructure operators: rural hospitals, community banks, local utilities. The Treasury, alongside CISA and the renamed Department of War, would have 60 days to build a classified benchmarking process for evaluating advanced AI models. The Attorney General would be told to prioritise existing criminal statutes, particularly the Computer Fraud and Abuse Act, against anyone using AI to break into systems.
None of that is novel in principle. The architecture in the middle, though, is. The draft creates a category called a "covered frontier model" and a voluntary process in which developers could hand the federal government access to such a model "for a period of up to 90 days before they plan to release such models to other trusted partners." That window would let the NSA and CISA probe the system for cyber capabilities before it ships, in exchange for confidentiality and nondisclosure protections.
It is the 90-day window that appears to have killed the order. According to Politico and Axios, Elon Musk, Mark Zuckerberg and David Sacks pushed back against the draft, with concerns reportedly centred on the delay it could create for product timelines and for sharing models with allied governments running their own evaluations. OpenAI, by contrast, was reported to support the order. Trump told Axios he did not want to "do anything to get in the way" of US dominance over China.
And yet the draft goes to unusual lengths to insist on what it is not. "Nothing in this section," it says, "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." That line, more than the 30-day deadlines or the Treasury's curious role in benchmarking AI capabilities, sits at the centre of the political problem. It tries to build a review apparatus while explicitly foreclosing any interpretation of it as a license. The administration evidently believed both halves could coexist. The companies that read the draft did not.
The choice of Treasury is also worth noting. The draft does not explain why a department whose closest brush with AI evaluation is sanctions enforcement should help set the benchmarks for what counts as a covered frontier model, and observers flagged it as one of the details that drew the most internal scrutiny. The presence of Treasury, alongside the NSA and CISA, hints that the framework was built less around technical evaluation than around classification authority and information control.
None of this exists in a vacuum. The same administration in February ordered federal agencies to stop using Anthropic's technology over a dispute about Claude's usage policies on mass surveillance and autonomous weapons. The supply chain risk designation that followed has been in and out of the courts ever since, with oral arguments in a D.C. Circuit case heard just three days before the planned signing. The shelved order would have created a different and softer channel for the government to look inside commercial AI systems. Whether that softness would have eased the existing fights or amplified them is impossible to say from a draft that was never signed.
For now, the deadlines run from no date at all. The order document is dated "May XX, 2026." Trump's wording suggests revision rather than abandonment, and the underlying questions, about who evaluates frontier models, on what timeline, under what classification, are not going away. The next draft, when it appears, will have already shown its hand.