The invitations were printed. The signing ceremony was scheduled. AI executives had booked flights to Washington. Then, between Wednesday night and Thursday morning on May 21, three men made phone calls to the President of the United States, and the most substantive piece of federal AI oversight ever drafted by this administration disappeared.
Elon Musk, Mark Zuckerberg, and David Sacks — Trump’s outgoing AI and crypto adviser — spoke directly with the president in the hours before the signing. By Thursday morning, the executive order was dead.
What the Order Would Have Done
The scrapped executive order was not the regulatory crackdown that Silicon Valley likes to warn about. It was a voluntary framework — no licensing requirements, no mandatory hold periods, no new federal agency.
Here’s what was actually on the table:
- Voluntary pre-release testing: AI companies could submit advanced models to federal agencies for security review up to 90 days before public release. The emphasis is on “could” — participation was optional.
- Treasury-led vulnerability scanning: The Treasury secretary would have led a voluntary partnership with industry to scan AI-integrated software for security flaws.
- CISA cybersecurity upgrades: Binding directives for civilian government agencies to upgrade their own cybersecurity against AI-enabled threats.
- AI crime prosecution: The Attorney General would have been directed to prioritize prosecuting AI-enabled computer crimes.
- OMB grant review: The Office of Management and Budget would have assessed redirecting existing grant programs toward AI vulnerability detection.
Six agencies were involved in drafting: Treasury, the NSA, CISA, OMB, the Office of Personnel Management, and the Department of Defense. The classified benchmarking process would have given the government its first structured look at frontier models before they ship to hundreds of millions of users.
The order was, by any measure, modest. It asked AI companies to voluntarily let the government check their homework. They said no.
Who Killed It and How
The Washington Post reported that Sacks, Musk, and Zuckerberg all communicated with Trump between Wednesday night and Thursday morning. The argument that landed, according to Axios, was an appeal to the “accelerationist” faction — officials at the National Economic Council who believe any oversight, even voluntary oversight, risks slowing the U.S. down in the AI race against China.
Sources described the order as “something doomers wanted,” framing even optional safety testing as an ideological position rather than a practical measure.
Trump’s public explanation was blunt: “I think it gets in the way of — you know, we’re leading China, we’re leading everybody, and I didn’t want to do anything to get in the way of that lead.”
The denials came quickly. Musk posted on X: “This is false. I still don’t know what was in that executive order and the President only spoke to me after declining to sign.” Meta said Zuckerberg only spoke with Trump after the decision was already made. Both claims are difficult to verify and contradicted by reporting from multiple outlets citing White House sources.
The Accelerationist Win
What happened here goes beyond one executive order. The scrapping represents a clear victory for the accelerationist wing of the administration over anyone advocating for AI governance, however minimal.
Consider what was rejected: a voluntary program. Not a mandate, not a licensing regime, not a regulatory body with enforcement power. The government asked to be allowed to look at AI models before they’re released to the public, and the industry’s most powerful figures convinced the president that even asking was too much.
OpenAI, notably, supported the order. The company has been pursuing state-level AI regulation strategies, reportedly with White House approval — an ironic position given the administration’s earlier executive order threatening states that adopt AI rules the federal government doesn’t like.
The result is a regulatory vacuum that is now deliberate rather than accidental. Europe has the AI Act. China has its own AI governance framework. The U.S. has a president who just publicly stated he won’t let oversight “get in the way.”
Why This Matters
The practical consequences are straightforward. When the next frontier model ships with a security vulnerability — and given recent history with prompt injection RCEs in Microsoft’s Semantic Kernel and authentication bypasses exploited within hours of disclosure — there will be no government review process that could have caught it.
When an AI model generates outputs that enable real harm — a credible bioweapons synthesis pathway, a novel cyberattack technique, a convincing impersonation used in fraud — there will be no pre-release checkpoint where someone outside the company could have flagged the risk.
This isn’t hypothetical caution. The Verizon 2026 Data Breach Investigations Report found that vulnerability exploitation now accounts for 31% of all breaches, surpassing credential abuse for the first time. Mandiant’s M-Trends report found that 28.3% of vulnerabilities are exploited within 24 hours of disclosure. The attack surface is growing faster than the industry’s ability to secure it, and the government just chose to look away.
Meanwhile, the three men who made those phone calls run companies — xAI, Meta, and various Sacks-backed ventures — that stand to benefit directly from releasing AI products without government review. This isn’t a conspiracy; it’s a straightforward conflict of interest that nobody in the administration appears concerned about.
What Comes Next
The White House Office of the National Cyber Director has indicated it’s working on “alternative AI security initiatives,” but no timeline or scope has been announced. In practice, this likely means the executive order is dead rather than delayed.
The U.S. approach to AI governance is now clear: there won’t be one. Not because Congress tried and failed to pass legislation, but because three billionaires picked up the phone and convinced the president that even asking nicely was a threat to American competitiveness.
For users of AI products — which, at this point, is most people with a smartphone — this means the safety testing of the models you interact with daily is entirely at the discretion of the companies that built them. Some of those companies take safety seriously. Others have already shown they’ll ship fast and patch later.
The question is whether you trust every AI company to police itself. If the answer is no, there is currently no backup plan.