LLM Guard vs NVIDIA NeMo Guardrails

LLM Guard is a self-hosted security toolkit for checking interactions with large language models. It scans prompts and outputs for concerns such as prompt injection, secrets, personally identifiable information, toxicity, bias, malicious URLs, and factual consistency. It can sanitize interactions, detect harmful language, help prevent data leakage, and resist prompt injection attacks. The documentation includes examples for the OpenAI SDK, LangChain, LiteLLM, and LlamaIndex. In the OpenAI examples, prompts and responses are scanned; a streaming example also checks response chunks. LLM Guard can run as an API, and custom scanners can be added. Installation uses `pip install llm-guard` and requires Python 3.9 or higher. It is free and listed under the MIT License, with API, Linux, and self-hosted platforms. The repository was archived on July 9, 2026, is read-only, and is no longer actively developed or maintained. Production deployment assistance is directed to GitHub issues, Slack, and [email protected].

NVIDIA NeMo Guardrails is an open-source Python package for adding programmable policy and safety checks to applications built with large language models. It can block, alter, or validate inputs and model responses that are unsafe, off-topic, malicious, or against configured policies. Its rails cover input, retrieval, dialog, execution, and output stages. Developers can configure behavior with YAML files, Colang flows, built-in guardrails, and custom actions. The catalog includes content safety, jailbreak protection, topic control, PII detection, agentic security, tool calling, hallucination and fact-checking, and model self-checks. Applications can use the Python SDK or a Guardrails server with HTTP endpoints such as OpenAI-style chat completions. The library supports Windows, Linux, and macOS, and can run in a self-managed Python environment. The installation guide lists Python 3.10 through 3.13 and at least one CPU with 4GB RAM; external models may require separately deployed GPUs. The open-source library costs nothing, though external models and services may need separate setup. NVIDIA cautions that no single rail covers every adversarial prompt and advises defense in depth and evaluation.