GPT-OSS-Safeguard (openai.com)

OpenAI released gpt-oss-safeguard, a research-preview family of open-weight reasoning models for safety classification (gpt-oss-safeguard-120b and -20b) under the Apache 2.0 license, downloadable from Hugging Face. Unlike traditional classifiers that learn a fixed decision boundary from labeled examples, these models take a developer-provided policy at inference time and output a classification plus chain-of-thought reasoning. That “bring-your-own-policy” design makes policy iteration fast, supports nuanced or emerging harms, and helps teams with limited labeled data produce explainable labels — useful when latency is less critical than interpretability. Technically, gpt-oss-safeguard was trained with reinforcement fine-tuning to mirror expert judgments (the Safety Reasoner approach) and excels at multi-policy accuracy in internal evaluations, outperforming gpt-5-thinking and prior open models on that metric. Public-benchmark results are mixed: slight wins on OpenAI’s 2022 moderation set, but modestly behind on ToxicChat. Limitations include higher compute/latency costs and cases where large, dedicated classifiers trained on tens of thousands of examples still outperform policy-driven reasoning. OpenAI is launching this preview with ROOST and a model community to solicit feedback and documentation; the release makes a production-grade, explainable safety tool available for anyone to adopt, modify, and integrate into multi-layered moderation stacks.