Scoop: Inside Trump's AI framework

The Trump administration has announced a framework to evaluate advanced AI models before their release, but the White House has decided not to make it public. Sources familiar with the matter tell Axios that the framework, reviewed on Tuesday, sets a specific definition for "frontier models" - those that are closed-source, possess state-of-the-art capabilities, and pose national security risks.

However, the criteria for what constitutes state-of-the-art or a national security risk remain unclear. Open-source models are specifically excluded from this framework.

Under the proposed 30-day pre-release review period, government employees would be barred from accessing these models, which would instead be stored in highly secure environments. Detailed logs would be required to track who accesses these models during the review process.

The review will involve multiple administration officials rather than a single office or agency. Despite the voluntary nature of the framework, the White House has chosen not to disclose its contents publicly. Meetings with industry leaders were held at the White House to discuss the framework, but those not invited are left in the dark about its specifics. It is also uncertain which "trusted partners" will have early access to these advanced models, including any foreign governments that might qualify.

The executive order mandates that the benchmarking process to assess AI models' cyber capabilities be classified, and there is no requirement to publicly release the voluntary framework under this order. Companies are advised to share models nearing public release rather than early stage ones in response to these discussions. However, many companies with less advanced systems may not qualify for the framework, according to sources.