US to Stress-Test Advanced AI Models

U.S. to Safety Test New AI Models The U.S. Department of Commerce will now safety test new artificial intelligence (AI) models from Google, Micrososources, and xAI before their public release. This initiative, announced recently, expands existing agreements with AI developers and aims to evaluate AI models for capabilities and security through Commerce's Center for AI Standards and Innovation (CAISI).

This voluntary submission process builds upon prior agreements with companies like OpenAI and Anthropic. CAISI has previously conducted 40 evaluations, including unreleased state-of-the-art models. The expanded collaboration will involve testing, collaborative research, and best practice development for commercial AI systems, ensuring public interest and national security considerations are addressed.

The move signifies an increased focus on AI oversight by the U.S. government, particularly as AI tools like Google's Gemini are integrated into defense agencies and concerns arise over the power of advanced models. Micrososources has affirmed the necessity of government collaboration for national security and public safety risks associated with large-scale AI.