The article outlines prompt injection as a principal adversarial AI threat, detailing how it involves manipulating text prompts in large language model (LLM) systems to elicit unauthorized actions. It underscores the fundamental design flaws in models that do not distinguish between legitimate instructions and user-injected prompts. The piece further classifies prompt injection into direct and indirect types and highlights its exploitation in multimodal GenAI systems, including image-based prompts. The range of potential attacks from prompt injection is vast, encompassing the exposure of sensitive information, control overrides, and data exfiltration. It contrasts the challenge of addressing these threats with traditional structured language injection attacks and concludes with a discussion on the early stages of LLM firewall approaches and the pressing need for innovative solutions in cybersecurity for GenAI.