Seeing the Threat: Vulnerabilities in Vision-Language Models to Adversarial Attack