White House Completes TRAINS AI Review Framework, and Keeps It Secret

The White House says it completed TRAINS, the voluntary frontier AI review framework ordered in June, by the August 1 deadline, and it will not publish the benchmarks that determine which models count as covered.

The program runs out of NIST and involves more than 10 federal agencies and 10 national labs. OpenAI, Anthropic, Google, Microsoft, and xAI have signed on. Meta says it expects to join soon but has not yet signed.

How TRAINS is supposed to work

President Trump’s June 2 executive order gave agencies 60 days to build a classified process for identifying the most dangerous models and a voluntary path for companies to submit those models for government review before release. Participating companies can face a roughly 30-day evaluation window tied to federal funding eligibility, including Defense Department AI budgets.

White House officials describe the secrecy as intentional: the thresholds measure a model’s ability to find and exploit software vulnerabilities, and publishing cutoffs could help adversaries tune models to stay just under the line. The framework is expressly non-mandatory, the order bars using it as a licensing or preclearance requirement.

Who wrote the rules, and who can’t read them

Critics note that the five participating labs helped shape covered-model criteria while smaller labs and open-source developers had no seat at the table. There is no Federal Register filing, and outside researchers cannot audit criteria they cannot see. If the NSA designates a model as a covered frontier model, companies cannot easily challenge thresholds that remain classified.

Open-weight models create a structural mismatch. Meta’s strategy of releasing downloadable weights conflicts with a 30-day pre-release hold, and the White House has not publicly explained how open-source releases fit the design.

Decoded Take

TRAINS is Washington’s bet that voluntary secrecy can govern frontier AI faster than legislation. The political tradeoff is sharp: speed and national-security opacity on one side; public accountability and open-source compatibility on the other.

The deeper story is dual regulation. While the U.S. builds a closed voluntary club for frontier labs, the EU is activating inspect-and-fine powers with public statutes. Global model providers now face two incompatible philosophies of oversight. Watch Meta’s eventual TRAINS agreement language closely, it will reveal whether open-weight AI can coexist with pre-release government review, or whether the framework quietly privileges closed models.