[2406.13294] Enhancing Cross-Prompt Transferability in Vision-Language Models through Contextual Injection of Target Tokens