The earlier modules showed how AI systems are attacked and where their weak points sit. This module is about the other side: the controls you put in place so that an attack, or a plain mistake, does not turn into real harm. No single control is enough, so the goal is defense in depth.
Guardrails check what goes into the model and what comes out of it, blocking prompt injection, leaked secrets, personal data, and unsafe content. Tools such as NeMo Guardrails and Llama Guard make this practical. Around them sit broader controls: least privilege for tools and agents, allow-lists for actions, human approval for risky steps, and monitoring that shows when something goes wrong.
Watch the lessons in order, then continue on to AI Risk Management Frameworks, where these controls are tied to the structured approaches organizations use to manage AI risk.
A beginner-friendly introduction to adding guardrails to an LLM application, using NVIDIA NeMo Guardrails, Llama Guard, and Guardrails AI.
A hands-on security lab that builds NeMo Guardrails around a locally hosted LLM to see how well it holds up against prompt injection.
A demonstration of Llama Guard, a safeguard model used to flag unsafe or policy-violating content in prompts and responses.
Shows how a multi-layered strategy works in practice, combining an AI security gateway with programmable guardrails to protect a GenAI stack.