What This Stack Does
As LLM applications move to production, security cannot be an afterthought. This stack combines five specialized tools that protect different layers of the AI security surface. NeMo Guardrails from NVIDIA provides programmable guardrails that control LLM behavior through dialog management, preventing off-topic responses, hallucinations, and unauthorized actions. LLM Guard adds input/output scanning for prompt injection, sensitive data leakage, and toxic content.
Vulnerability Scanning and Prompt Fuzzing
Garak functions as a vulnerability scanner specifically for LLMs, probing models for weaknesses like jailbreaks, prompt leaking, and data extraction. ps-fuzz focuses specifically on prompt injection testing, fuzzing your LLM endpoints to discover injection vectors before attackers do. ModelScan checks serialized ML models for malicious payloads — a growing attack vector as teams share and deploy pre-trained models.
The Bottom Line
Deploy NeMo Guardrails as the runtime safety layer, LLM Guard for real-time input/output filtering, and use Garak and ps-fuzz during development and CI/CD for pre-deployment security testing. ModelScan should run on every model artifact before deployment. This layered approach ensures your LLM application is protected at input, processing, model, and output stages.