Frontier AI safety tests may be creating the very risks they're meant to stop The Register