Defense-to-Attack Bypassing Weak Defenses Enables Stronger Jailbreaks in Vision-Language Models