Granite Guardian

Granite Guardian is IBM's family of open-weight safety models designed to sit in front of or behind any LLM and screen content against a defined risk taxonomy. The risks it targets are mapped to IBM's AI Risk Atlas and include hate, abuse and profanity, jailbreak attempts, RAG and tool-use hallucinations, and custom user-defined criteria, so teams can enforce their own policies rather than a fixed blacklist.

The line has progressed through several generations: Guardian 3.0 launched alongside Granite 3.0 in October 2024 (era-level date used here), followed by 3.3 and then 4.1, which adds hybrid thinking modes and reportedly ranks in the top three on the LLM-AggreFact hallucination-detection benchmark. Weights are released on Hugging Face under Apache 2.0, making Guardian one of the few fully open, commercially usable guardrail model families.