OpenAI drops open-weight reasoning models for safety

OpenAI just released a research preview of two open-weight models built specifically for safety classification, and they’re taking a fundamentally different approach than the industry standard. The models—gpt-oss-safeguard-120b and gpt-oss-safeguard-20b—are fine-tuned versions of the company’s gpt-oss open models and come under the same permissive Apache 2.0 license. You can grab both from Hugging Face right now.

The real shift here is philosophical. Traditional safety classifiers are trained on thousands of manually labeled examples, trying to reverse-engineer a policy they’ve never actually seen. OpenAI’s new models flip that entirely. A developer provides the policy directly at inference time, and the model uses chain-of-thought reasoning to interpret it and classify content accordingly. The policy lives outside the model, so developers can tweak it iteratively without retraining anything. That’s a genuine departure from the usual grind of curating massive labeled datasets every time a policy changes.

OpenAI sees this performing best in messy, real-world scenarios where harms are evolving fast, the domain is too nuanced for smaller classifiers, or a platform simply doesn’t have enough samples to train a bespoke model for every risk. They’re positioning it for cases where latency takes a backseat to explainable, high-quality labels. A gaming forum wanting to flag cheating discussions or a reviews site screening for fakes—those are the kinds of use cases where dropping in a custom policy on the fly makes sense.

The release is explicitly a preview meant to gather feedback from the safety and research community. OpenAI worked with ROOST, a nonprofit focused on online safety, to test the model and produce developer documentation. ROOST is also launching a model community today to explore open AI for protecting online spaces. Whether this reasoning-based approach holds up under real-world pressure is an open question, but it’s a practical acknowledgment that one-size-fits-all safety policies never really fit anyone well.