AI safety research should be conducted independently of commercial frontier labs