In this practical session, participants will learn how to build autonomous AI agents using open-source LLMs and apply responsible AI principles through real-world guardrailing techniques. We will walk through the full pipeline — from creating a task-specific agent using LLaMA or Mistral-based models, to integrating NVIDIA NeMo Guardrails, Llama Guard, and prompt-based safety strategies.

We’ll cover critical safety challenges such as:

• Prompt injection and jailbreaks
• Toxicity and bias mitigation
• Controlling agent autonomy and output

This session will include:

• Setting up an AI agent for a real-world use case (e.g., customer support, knowledge
  assistant)
• Injecting common adversarial prompts to test vulnerabilities
• Applying NVIDIA NeMo Guardrails and Llama Guard to detect and prevent harmful
  outputs
• Using prompt-based guardrailing as a first line of defense
• Discussing practical limitations and failure cases in alignment and safety