OpenAI plans to expand independent safety assessments into model training and evaluation, giving outside groups earlier access to test high-risk AI systems.