Which statement best describes prompt injection in AI systems?

• A. Prompt injection manipulates the model by altering the user-provided context or prompts to override safety constraints, causing outputs that violate guidelines.
• B. Prompt injection is a routine calibration method to improve safety.
• C. Prompt injection manipulates the model by altering system prompts to improve responses.
• D. Prompt injection manipulates the model by altering the user-provided context or prompts to override safety constraints, causing outputs that violate guidelines.

Answer: (A)

Prompt injection works by slipping instructions into the prompt itself that steer the model to ignore safety rules or produce outputs it normally wouldn’t. Because the model treats the user-provided prompt as its guiding instruction, adding phrases like “ignore safety rules” or other harmful directives can override safeguards and lead to unsafe responses. This focus on manipulating what the user submits distinguishes prompt injection from routine safety calibration, which is a legitimate, non-hraudulent process, and from attempts to adjust system prompts for better responses, which isn’t about exploiting the user’s input to bypass constraints. Understanding this helps you see why the user-provided prompt is the target and why the goal is to prevent injected instructions from steering the model away from its safety guidelines.