Guardrail frameworks wrap a model or agent with input checks, output checks and conversation rules defined in config or code. They are useful for consistent content and topic policy, jailbreak screening and tool call checks across many apps. Measure their false positive rate and latency, and remember they share the model's blind spots when an LLM is the judge.
Read in this order
- OfficialNVIDIA NeMo Guardrails library developer guideStart here: a full framework guide with rails defined in config.
- Open sourceLlamaFirewall (Meta PurpleLlama)An agent-focused guardrail you can run, with code scanning and alignment checks.
- PaperLlamaFirewall: An open source guardrail system for building secure AI agentsThe LlamaFirewall paper, with its design and measured results.