“Releasing AI models in open source is the safest approach because it enables a wider community to build both AI systems and the necessary safety and protection systems.”