President Trump's June 2 executive order on AI gave federal agencies 60 days to build two things: a classified process for identifying the most dangerous AI models, and a voluntary framework for letting companies submit those models for government review before release. The deadline was August 1. The White House says it hit both deliverables. The review program, called TRAINS, runs out of NIST and pulls in more than 10 federal agencies and 10 national labs. OpenAI, Anthropic, Google, Microsoft, and xAI have all signed on; Meta says it expects to sign an agreement soon but hasn't yet. On Tuesday, White House staff are scheduled to brief those companies on what the completed framework actually says — the first time most of them will see it.

1. Voluntary Works, and Secrets Aren't Sabotage (White House officials, National Cyber Director Sean Cairncross)

The framework is a security tool, not a regulation, and classifying the benchmarks is the point.

The framework is complete, at least officially. A senior White House official told The Next Web: "The voluntary framework outlined in the June 2nd executive order was complete by the deadline." On why it won't release the contents: "Just because things are unclassified that doesn't mean we are going to broadcast them to everyone."

Classifying the benchmarks is a deliberate security choice. The thresholds measure a model's ability to identify and exploit software vulnerabilities. Publishing the specific cutoffs would help bad actors tune their models to stay just under the line.

The framework is expressly non-mandatory. The executive order bars using it as "a mandatory governmental licensing, preclearance, or permitting requirement." The government can't force anyone to submit — the five companies in TRAINS signed up voluntarily. National Cyber Director Sean Cairncross has framed the AI security push as broadly inclusive: "We cannot achieve the president's vision without securing and bolstering our open-source ecosystem."

2. But Critics Can't Audit What They Can't Read (Sen. Maria Cantwell, AI safety researchers, smaller AI labs)

Secret rules aren't accountability — and the labs writing the criteria run through those same tests.

Five companies helped write the thresholds their rivals now have to clear. OpenAI, Anthropic, Google, Microsoft, and xAI participated in shaping TRAINS' covered-model criteria. Smaller labs and open-source developers had no seat at that table.

There's no public trail at all. NIST, CISA, and OSTP haven't published or announced the completed framework, and it hasn't been submitted to the Federal Register. Nothing in the public record holds it accountable. Sen. Maria Cantwell (D-WA), Senate Commerce Committee Ranking Member, has called for an approach "grounded in standards, testing and public infrastructure investment" rather than the administration's closed-door model.

Nobody outside the NSA can challenge a covered-model designation. If the NSA designates a company's model as a "covered frontier model," the company can't dispute criteria it's never seen. The NSA director holds sole authority over the thresholds, sharing them only "as appropriate."

3. Especially Congress and the States, Which Are Losing Their Say (Sen. Chris Murphy, 36 state attorneys general, Sen. Mark Warner)

No federal AI law governs actual AI harms yet — but the administration is trying to stop anyone else from filling that gap.

The framework comes packaged with a push to kill state AI regulation. A bipartisan House draft — "The Great American Artificial Intelligence Act" — proposes a three-year freeze on state laws governing AI model development, targeting California's training-data transparency law and watermarking requirements. Sen. Mark Warner (D-VA) called the White House's AI governance approach "several pages of broad goals" that lacks real enforcement and inappropriately preempts state authority — and said the administration has been "killing Senate Intelligence Committee provisions two years running" on foreign AI threats.

States already have active laws that no federal rule has touched. California, Texas, and Colorado all have AI laws in force: disclosure requirements, self-harm interventions, medical authority limits. No federal statute has preempted any of them. Rep. Valerie Foushee said the White House's AI approach "lacks meaningful guardrails and ignores the real impacts AI is already having on jobs, communities, and local resources." Rep. Josh Gottheimer said it "broadly blocks state AI laws and lacks key consumer protections around AI models and agent safety."

The GUARDRAILS Act is the direct counter. Sen. Chris Murphy (D-CT) and Sen. Brian Schatz (D-HI) introduced it in April 2026 to repeal Trump's state-law moratorium and bar federal funds from being withheld from states that act on AI. Murphy: "States have to be able to protect American families from a technology that could destroy millions of jobs, addict kids to harmful chatbots, and drive up utility bills." Thirty-six state attorneys general have formally opposed broad federal preemption; more than 200 state lawmakers have called on Congress to reject it. Brad Carson of Americans for Responsible Innovation called the bipartisan preemption proposal a "generational mistake."

4. And Open-Source AI Still Doesn't Have a Clear Path In (Meta)

A 30-day pre-release window and open model weights don't fit together — and no one has solved that yet.

All major closed-weight AI labs are in TRAINS now — except Meta. A Meta spokesperson said the company "shares the goal of advancing US leadership on robust and secure frontier AI and expects to sign an agreement soon."

The 30-day hold doesn't fit open-source AI. Meta plans to sign anyway, it says — but Meta's strategy is to release model weights openly so anyone can download and run them. Holding those weights back for a 30-day government review window doesn't fit that model — and the White House hasn't publicly explained how open-weight models fit the framework at all. As one analyst put it: "It's hard to give the government a 30-day head start on evaluating your model when your entire strategy involves publishing the weights for the world to see."

Pressure to join is coming from the top. The White House has pushed Meta directly to submit its models for pre-release testing, with no reported consequences for refusal. The White House invited Meta to Tuesday's briefing.

Where This Lands

The administration is betting that voluntary reviews with classified benchmarks can govern AI risks without stifling innovation — and that the companies best positioned to help design those benchmarks are the ones already building the most powerful models. Critics say no one can hold accountable a framework they can't read, and that the labs writing the criteria are the same ones running through those tests. Congress and the states say the real fight isn't the framework at all: it's whether Washington gets to decide that nobody else can regulate AI. Meta's open-source strategy sits outside the current design, and Meta's planned participation doesn't resolve how open-weight models fit a system built around pre-release holds.

Sources