AI Security, Safety, and Privacy

Introduction to AI Security, Safety, and Privacy

AI security protects the system and its data from deliberate attack. Safety reduces harmful behavior and unintended consequences. Privacy limits how personal or sensitive information is collected, exposed, and retained.

An LLM system can be attacked before training, during inference, or through the software connected to the model. The first task is to identify which stage and asset an attack reaches.

Draw the complete surface: training data, model parameters, prompts, generated outputs, retrieval sources, and connected tools. A defense that protects only the chat prompt leaves the other paths untouched.