White House finishes classified-by-default AI evaluation framework, publishes nothing

The White House says it hit the 2 June executive order deadline for a voluntary framework to evaluate advanced AI models, but will not disclose its contents, who has read it, or when companies begin using it — and one official said unclassified does not mean it gets broadcast. Under the program, developers could hand the government pre-release model access for up to 30 days, while both the cyber-capability benchmarking process and the threshold for which models are covered stay classified. For frontier labs, the compliance surface is now set by a document only the government controls access to, and outside researchers have no way to check whether the evaluations are rigorous.

Source: axios.com ↗