Is open-source AI safer or more dangerous than closed models?