Attack where a malicious user includes hidden instructions in their input to manipulate the model into ignoring its original instructions, revealing confidential information, or performing unauthorized actions.
In AIF-C01 (Domains 3 and 5) it's distinguished from jailbreaking (trying to evade the model's safety restrictions) and prompt poisoning (contaminating training data). Mitigation: Guardrails, input validation, and a clear separation between system instructions and user content. In GENAI-L (Domain 4) it's one of the security risks SAIF aims to mitigate in generative AI applications.
Prompt injection is a concept that can show up on the exam. Check the full guide to see it in context.