Llama Guard is a family of open models from Meta that check the prompts going into, and the answers coming out of, an AI system for unsafe content.
In plain words
Llama Guard is itself a language model, fine-tuned to act as a safety classifier rather than a chatbot. You place it beside your main model: it reads a user's message or the model's reply and labels it safe or unsafe against a list of hazard categories. Llama Guard 4 is built from Llama 4 Scout, follows the MLCommons hazards list, and can check text in several languages as well as prompts with images.
A quick example
Imagine a chatbot company that runs every user message through Llama Guard before its main model sees it, and every reply through it again before the user does. Anything labelled unsafe is blocked or replaced with a refusal.
Where you'll meet it
- Input and output filtering for chatbots.
- Moderation around open-weight models such as Llama.
Related
Sources
- DocsLlama Guard 4 12B model card, Meta (Hugging Face) · read 29 Sept 2026
- DocsPurple Llama README, Meta (GitHub) · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.