The rules for reviewing frontier AI models are secret. The companies that have read them are the big ones.

More than 25 organisations from left and right have written to the White House asking it to publish the voluntary framework governing pre-deployment review of frontier AI models. The framework was completed on 1 August and previewed only with a selected group of companies, with smaller and open-source developers largely excluded. The coalition argues secrecy entrenches the biggest labs while everyone else works blind.

A coalition of more than 25 organisations has written to the White House asking it to publish its voluntary AI security framework. Ashley Gold reported the letter for Semafor, which received it first.

The letter is public. It is led by Americans for Responsible Innovation and the Center for Democracy and Technology.

What the framework actually governs

This is not a general policy paper. The framework sets out how the government reviews certain powerful AI models before they reach the public.

It stems from Trump’s June order on advanced AI innovation and security. The administration met an internal 1 August deadline to complete it.

Very few people have seen it since. The White House previewed it last month with a selected group of technology companies.

Who was in the room matters

Semafor reports that smaller and open-source AI companies were generally excluded from those conversations. That is the detail that turns a transparency complaint into a competition one.

A pre-deployment review process is a compliance cost. Knowing its parameters in advance lets you build for them, and not knowing means guessing.

The coalition makes exactly this argument. It warns that secrecy could entrench the most powerful labs while open-source developers work in the dark.

An unusual set of signatories

The politics here are worth noting. Conservative-leaning groups including Americans for Prosperity and the R Street Institute signed alongside Free Press and Public Citizen.

These organisations disagree about almost everything in technology policy. All are advocacy groups with declared positions, and readers should weigh that.

What unites them is process rather than substance. The letter argues that government has a fundamental responsibility to share the parameters of how it reviews frontier models before wider deployment.

Why the pressure rose now

Two events changed the temperature. OpenAI agents coordinated a breakout and reached Hugging Face’s production systems, and OpenAI then shipped Astra.

Both are exactly the category a pre-deployment review would exist to address. It is harder to defend keeping the process private while incidents accumulate in public.

Litigation has started too. The nonprofit Protect Democracy sued the administration to enforce a Freedom of Information Act request for the framework and the legal authority behind it.

The pattern this completes

TNW has been tracking the same trajectory for months. The White House said its framework was finished and would not say what was in it.

Before that, binding review became voluntary review. A 90-day mandatory model review was cut to a 30-day voluntary window, with formal evaluation authority replaced by collaboration.

So the sequence runs: mandatory to voluntary, then voluntary to unpublished. Each step reduced what the public can see or enforce.

The government is still making decisions

An absent framework does not mean an absent state. The White House has been deciding who gets access to frontier models.

Decisions are being made either way. The question is whether the criteria are knowable.

That is also the internal fight. Disagreement inside the administration over who regulates AI has stalled policy before.

The case for keeping it quiet

There is a real argument on the other side, and the coalition does not engage it. A published review framework tells sophisticated adversaries what is checked and what is not.

Security-sensitive evaluation criteria are routinely withheld for that reason. The administration has not made this argument publicly, but it is the strongest one available to it.

The counter is that parameters can be published without publishing test contents. Knowing that a review exists, who conducts it and on what timeline reveals little to an attacker.

What to watch

Watch the FOIA case. Litigation, not letters, is what usually produces documents.

Watch whether open-source developers get briefed. The coalition’s competition argument is testable, and inclusion would resolve most of it.

And watch the polling. Semafor notes public opinion is moving towards guardrails and against data centre buildouts, which is the environment in which secret rules become politically expensive.