“Open-source models increase AI safety by enabling broad access and understanding, whereas restricting them only harms positive actors.”