“Vision-language model (VLM) powered robotic systems are already being successfully jailbroken, demonstrating a vulnerability to prompt injection attacks.”