How to Secure LLMs with NVIDIA NeMo Guardrails | AI Security

Can NVIDIA NeMo Guardrails help protect a locally hosted LLM against prompt injection? This hands-on AI security lab demonstrates how to build an NVIDIA NeMo Guardrails security gateway from scratch and connect it to the Lily Uncensored Cybersecurity model running locally through LM Studio. The goal is to inspect requests before they reach the model and check responses before they return to the user. This video covers the environment setup, gateway deployment, and initial controlled prompt-injection tests. WHAT YOU’LL SEE • Introduction to NVIDIA NeMo Guardrails • How input and output guardrails work • Building a local AI security lab • Using Parrot OS as the testing machine • Running NeMo Guardrails in a Windows VM • Hosting the local LLM through LM Studio • Testing network connectivity between systems • Installing Python 3.12 from PowerShell • Creating a Python virtual environment • Installing NVIDIA NeMo Guardrails with pip • Connecting NeMo Guardrails to an authenticated LM Studio API • Creating input and output self-check policies • Building a FastAPI security gateway • Configuring baseline, input-protected and fully protected modes • Starting the gateway and testing access from Parrot OS • Sending controlled prompt-injection requests • Comparing responses with and without guardrail enforcement • Testing a synthetic canary instead of real credentials LAB ARCHITECTURE Parrot OS → NeMo Guardrails Gateway → Local LLM through LM Studio Fully protected request flow: Parrot OS → Input Guardrail → Local LLM → Output Guardrail → Response MODEL REFUSAL VS GUARDRAIL BLOCK A language model may independently refuse a malicious request. That does not necessarily mean the guardrail stopped it. This lab examines that distinction to show where requests are rejected and how the security controls behave independently from the underlying model. TOOLS USED • NVIDIA NeMo Guardrails • LM Studio • Lily Cybersecurity 7B • Python 3.12 • FastAPI and Uvicorn • Parrot OS • PowerShell • curl and jq RESOURCES NVIDIA NeMo Guardrails — GitHub: https://github.com/NVIDIA-NeMo/Guardr... NVIDIA NeMo Guardrails — PyPI: https://pypi.org/project/nemoguardrails/ 🤝 INVEST IN A SAFER FUTURE FOR AI Kiso Secure is an early-stage AI security company based in Kenya, focused on helping teams identify and address security weaknesses in models, AI agents and RAG applications. Its services focus on attack surface mapping, adversarial testing, reporting and remediation, with governance and compliance readiness. These practical labs offer a look at the technical work behind Kiso Secure’s mission. The company is seeking investors who share a commitment to privacy, meaningful human oversight and responsible AI development. Explore the company profiles and book an introduction to discuss Kiso Secure’s direction and investment opportunity. SeedForge: https://seedforge.com/s/n8rtxeHANrKA Gust: https://gust.com/companies/kiso-secure 📅 Book an investor introduction call: https://calendly.com/kisosecure/kiso-... CONNECT WITH KISO SECURE Follow for practical AI security labs, AI red teaming, LLM security testing, AI safety, governance and compliance. 🧡 Patreon — Support Kiso Secure:   / kisosecure   💼 LinkedIn — Company Page:   / kiso-secure   ▶️ YouTube:    / @kisosecure   💬 X: https://x.com/KisoSecure 📸 Instagram:   / kisosecure   🎵 TikTok:   / kiso.secure.ai   RESPONSIBLE TESTING This lab is designed for authorized AI security research and defensive testing. Use synthetic credentials, tokens, and test data when evaluating prompt injection or sensitive-information disclosure. #NVIDIA #NVIDIANeMo #NeMoGuardrails #AISecurity #LLMSecurity #PromptInjection #AIRedTeaming #LMStudio #CyberSecurity