The White House has an AI safety test for frontier models. It won't show anyone.

The Trump administration finalized its framework this week for testing the most powerful AI models on safety and cybersecurity risks, and the framework is not for public release. Testing criteria will be shared only with the small group of companies whose models are being vetted, according to The Guardian, CNBC, Axios, and NPR.

Under the June executive order that set the process in motion, the benchmarks and thresholds the government will use to grade frontier AI systems are classified. The companies submit their strongest models for review up to 30 days before public release, on a voluntary basis, and the results of those tests stay inside the room.

A meeting on Tuesday brought together staff from OpenAI, Anthropic, Meta, Google, Nvidia, and Microsoft to walk through the framework. OpenAI and Anthropic declined to comment. The voluntary character of the process is set out in the order itself: Executive Order 14409, signed in June, says nothing in the order authorizes mandatory preclearance for any AI model.

The order grew out of an April scare. Anthropic held a model it calls "Mythos" back from release because, the company said, the system could be used to break into corporate IT and financial networks. The episode briefly destabilized the administration's posture on AI oversight, according to The Guardian's reporting, and pushed Washington toward a more formal review process. By June, the order named Treasury, the National Security Agency, and the Cybersecurity and Infrastructure Security Agency as the agencies responsible for building a classified benchmark, and set August as the deadline for the framework to land.

The deadline has been met, and the framework has been kept out of public view. Treasury, the NSA, and CISA declined to make the benchmark public. The participating companies have not explained what the tests measure, what scores would trigger a follow-up, or whether the results will ever be disclosed outside the companies whose models are reviewed.

The order was not the version of oversight that some in the administration originally wanted. The Guardian reports that Elon Musk and Mark Zuckerberg personally lobbied President Trump against an earlier, mandatory version of the vetting requirement, and that the June order is the watered-down result. The original proposal would have given the government a binding role in deciding which models could ship. The current version asks companies to submit voluntarily, and shares the grading rubric only with the submitter.

That sequence is what makes the final framework a private certification channel rather than a public review. The government classifies the test. The company hands over the model. The government returns a result. Outside researchers, foreign governments, smaller AI labs, and the businesses that depend on these models have no formal read on what was measured or what passed. NPR, in its coverage of the order, described it as a voluntary program that asks companies to submit their strongest models before release.

The next test is the next release. The first wave of voluntary submissions is expected before the end of the year, and the framework's effectiveness will be judged by what the government knows, what the labs know, and what the public learns. On current settings, the answer is in that order.