Will releasing the weights of future large language models grant widespread access to pandemic agents?