Rebedea et al. at NVIDIA published NeMo Guardrails, an open-source toolkit using Colang programmable rails to enforce topical boundaries, fact-checking, and safe tool execution around LLM apps.
ConceptLLMCross-Modal
Guardrails & AI Safety
Also known as: Guardrails · Llama Guard · NeMo Guardrails · Constitutional Classifiers
Plain English
Runtime safety checks, input/output classifiers, and programmable rules that wrap AI models to block prompt injections, jailbreaks, and policy violations in production.
Technical Overview
Production AI guardrails layer lightweight safety classifiers (Llama Guard), programmable dialogue rails (NVIDIA NeMo Guardrails), and streaming token-level Constitutional Classifiers around foundation models to enforce topic boundaries, PII redaction, and jailbreak defense alongside schema validation.
Timeline (3)
2023
2025
See also (bkz.) (9)
Open full graphGuardrails & AI Safety