Align to Misalign: Automatic LLM Jailbreak with Meta-Optimized LLM Judges