US Government Finalizes Voluntary AI Model Evaluation Framework Behind Closed Doors

The US government has completed a voluntary evaluation framework for advanced AI models, developed in collaboration with major AI labs including OpenAI, Anthropic, and Google. However, the White House has chosen not to disclose the specific details, thresholds, or participating entities of the framework to the public. This framework represents a significant step in US AI governance and safety regulation, establishing protocols for pre-release government access to powerful models. However, the lack of public transparency limits immediate external scrutiny and leaves broader policy observers in the dark about exact regulatory thresholds. The framework defines confidentiality, cybersecurity, and intellectual property protections for when the government accesses models up to 30 days before release. While the framework itself is not classified, the specific benchmarks for testing cyberattack capabilities and the model size thresholds remain classified.