Frontier-AI companies cannot be trusted to regulate their own security risks