How authoritarian governments twist AI safety to coerce tech companies to comply