Large Language Models (LLMs) have quickly improved in how well they understand and generate human-like language. But as they become more capable, they also become more vulnerable to adversarial manipulation. This survey looks at different types of prompt-based attacks that take advantage of the tendency of models to follow instructions, often in ways that can undermine safety, privacy, or reliability. We organize these threats into a clear taxonomy and also explore a range of defense strategies. In addition, we review tools and benchmarks used to test how robust these models are (including PyRIT, Giskard, Garak, and PromptBench). By mapping attacks to defenses in a layered framework, this work emphasizes the need for thoughtful, flexible safeguards when using LLMs in real-world settings. Content Warning: This paper contains examples of harmful language.