Secure Linear Alignment of Large Language Models