WASHINGTON: President Donald Trump is expected to sign an executive order on Thursday, May 21, 2026, that establishes the first federal framework for government review of frontier artificial intelligence models before public release, according to a draft of the order described by Bloomberg. The order pairs a voluntary pre-launch review program for so-called “covered frontier models” with a Treasury-led cybersecurity clearinghouse and expanded hiring inside the government’s AI engineering corps. Tech industry leaders have been invited to attend the signing event at the White House.
The order stops short of mandatory federal approval. AI labs that opt in would share advanced models with the government for review, on a timeline that remains under negotiation. One draft of the order sets the window at 90 days before public release. Industry participants, including OpenAI and Anthropic, have pushed for around 14 days, according to Axios, which first reported the framework on May 20.
Why the White House Reversed Course on Frontier AI
The order marks a meaningful shift from the deregulatory posture Trump set in his January 2025 executive order, “Removing Barriers to American Leadership in Artificial Intelligence,” and from EO 14365 in December 2025, which laid out the administration’s national AI policy framework. Both earlier orders treated frontier model oversight as a constraint on American competitiveness. The new order treats it as a national security necessity.
The catalyst was Anthropic’s April 2026 disclosure of Mythos, an internal model the company said could rapidly find and exploit cybersecurity vulnerabilities at a scale beyond what most human researchers can match. Anthropic declined to release Mythos publicly and instead launched Project Glasswing, a defensive program offering up to $100 million in usage credits to select cybersecurity firms. The disclosure ricocheted through Washington and is widely credited inside the White House as the trigger for moving the order to signing.
The pressure has been building for months. Anthropic’s profile inside Washington jumped again in May after the company closed a $30 billion round at a $900 billion valuation, briefly surpassing OpenAI on paper and turning every Anthropic product disclosure into a policy event. Separately, the rise of self-improving systems, including the recursive superintelligence work behind Richard Socher’s $650 million bet on Recursive Superintelligence, has pushed national security officials to treat any frontier release as potentially dual-use.
The cybersecurity section of the order revamps existing federal threat-sharing programs to cover AI companies and authorizes a clearinghouse, formed by the Treasury Department alongside other agencies, that would identify and remediate vulnerabilities in unreleased models. It also calls for expanded hiring at the U.S. Tech Force, the government’s in-house AI engineering team, and protections for cybersecurity infrastructure across hospitals, banks, and other operators of critical systems.
What does Trump’s AI executive order mean for founders building on frontier models?
For most founders, nothing changes overnight. The order targets companies operating at the frontier (OpenAI, Anthropic, Google, and a small set of peers), not the application layer. But anyone building on top of those APIs should expect occasional release delays as covered models move through review, plus new compliance language flowing downstream in enterprise contracts and data-sharing addenda.
The voluntary structure also matters. Because the program is opt-in, lab participation becomes a competitive signal. A startup that pre-clears its frontier model with the government can market that to enterprise and federal buyers in a way a non-participant cannot. Expect “reviewed under the frontier model framework” to start appearing in security questionnaires and procurement bids, much like SOC 2 attestations did a decade ago. Founders selling AI tooling into regulated industries should watch how Treasury defines “covered.” The threshold will determine which providers can credibly compete in financial services and healthcare contracts, and how policy headlines like Trump’s order shape the next round of enterprise AI pricing and packaging from the platforms founders depend on.
The review window itself is the live question. A 90-day window would meaningfully delay product roadmaps for any startup downstream of a covered model. A 14-day window, which OpenAI and Anthropic have argued for, would be closer to a standard security review and less disruptive to shipping. The final text, once published in the Federal Register, will settle it.
What to Watch Over the Next 90 Days
Three things will determine whether this order reshapes the AI market or fades into compliance theater. First, the published text. The 90-day versus 14-day disagreement is the headline negotiation, but the definition of “covered frontier model” matters more. A narrow definition limits the order to a handful of labs. A broad definition pulls in mid-tier providers and changes the calculus for any startup approaching the threshold.
Second, who signs up. Voluntary frameworks gain authority through participation. If OpenAI, Anthropic, Google, and a critical mass of frontier developers commit publicly at the signing, the program acquires de facto standard status. If participation is uneven, the framework becomes a checkbox that lab-by-lab compliance officers learn to manage around.
Third, what Treasury’s clearinghouse actually does. A vulnerability disclosure pipeline is only as useful as the agency running it. Treasury has cybersecurity experience through OFAC and FinCEN sanctions enforcement, but no track record on AI model security. Whether the clearinghouse hires the right engineers, and whether labs trust it with pre-release model weights, will decide if the program produces real security findings or stays symbolic. Expect the first public read on participation and clearinghouse operations within the next quarter.
