Frontier AI models should undergo mandatory third-party safety evaluations before release