Granite Guardian 3.3 8B

Granite Guardian 3.3 8B is IBM's open safety classifier for detecting risks in LLM prompts and responses, released in August 2025. It comes pre-baked with detectors for jailbreak attempts, profanity, and hallucinations in tool calls and retrieval-augmented generation, and supports bring-your-own-criteria so teams can define custom judging rules.

The 3.3 release adds a hybrid thinking mode that exposes reasoning traces alongside verdicts, and IBM reports it outperforming much larger models on fact-checking benchmarks such as LLM-AggreFact. It is released under Apache 2.0.