Basic Concept: Input manipulation attacks — including adversarial examples and prompt injection — alter inputs to cause AI systems to produce unintended, harmful, or inappropriate outputs. Defending against these attacks requires mechanisms that validate and constrain both inputs and outputs at runtime. CompTIA SecAI+ Study Guide identifies guardrails as the primary defense against input manipulation in AI systems.

Why D is Correct: Guardrails implement real-time validation and filtering of both incoming inputs and outgoing recommendations. They detect manipulated inputs that deviate from expected patterns, enforce content policies on outputs, and prevent the system from producing inappropriate recommendations regardless of how cleverly the input was crafted. Guardrails provide the most comprehensive and directly applicable defense against the described manipulation attack scenario.

Why A is Wrong: Cross-validation is a model evaluation technique that assesses how well a model generalizes to independent datasets during training. It measures predictive performance but does not provide runtime protection against input manipulation attacks on a deployed system.

Why B is Wrong: Feature regularization is a training technique that adds penalties to model weights to prevent overfitting. It improves generalization during training but does not inspect or validate inputs at inference time to detect manipulation.

Why C is Wrong: Feature scaling normalizes input feature values to a standard range for training efficiency. Like regularization, it is a preprocessing and training step that has no effect on defending against runtime input manipulation attacks on deployed systems.