Granite Guardian

Comprehensive risk coverage: Granite Guardian detects seven major risk categories including social bias, profanity, violence, sexual content, unethical behavior, jailbreaking attempts, and hallucination-related issues—addressing gaps that traditional safety models miss, particularly around jailbreaks and retrieval-augmented generation (RAG) problems.

Works with any LLM: Rather than being tied to a specific language model, these safeguard models function as an add-on layer that can protect users regardless of which LLM they're using, making responsible AI practices more universally accessible.

Strong performance metrics: The models achieved AUC scores of 0.871 for harmful content detection and 0.854 for RAG-hallucination detection, positioning them as the most generalizable risk detection models currently available according to the paper's benchmarking.

Trained on diverse, real-world data: The models were developed using both human annotations from multiple sources and synthetic data, helping them recognize nuanced risks that reflect actual usage patterns rather than theoretical scenarios.

Open-source for community benefit: By releasing Granite Guardian as open-source software, the developers aim to democratize responsible AI safeguards, enabling broader adoption of safety practices across organizations rather than concentrating these tools with a single company.