Prompt injection represents a new class of vulnerability in generative AI applications. Coined in 2022, it occurs when developers mix their own instructions with untrusted content, treating the model's response as if it were securely bounded. Unlike SQL injection, where there's a clear distinction between data and instructions, prompt injection lacks this boundary. Current large language models (LLMs) predict the next token without recognizing these differences, making them inherently vulnerable to manipulation. This vulnerability has emerged as a top concern for developers working with generative AI, as prompt injection attacks regularly compromise systems. SQL injection has been around for nearly 30 years and is well understood among security professionals. It allows attackers to execute harmful commands through seemingly innocent web forms. However, prompt injection operates differently. For instance, if a recruitment tool asks an LLM to evaluate CVs but a candidate embeds a hidden instruction to approve their CV, the model might execute that instruction as if it were valid input. This blurring of lines between data and instruction complicates mitigation efforts, as traditional methods like parameterized queries in SQL don't apply effectively to LLMs. To address prompt injection, developers need to shift their mindset. Rather than treating it as a code injection problem, it should be viewed as a "confused deputy" issue—where the model can be misled into executing harmful functions. Raising awareness among developers and security teams is essential, as many are not yet familiar with this vulnerability. Secure design practices must focus on deterministic safeguards to prevent misuse of the model’s outputs, particularly when interacting with external tools or APIs. The approach should center on risk reduction rather than expecting a complete mitigation of prompt injection vulnerabilities.