The gpt-oss-safeguard models use reasoning to interpret developer-defined policies during inference. This means developers can classify user messages and content according to their specific needs. Unlike traditional methods that rely on large datasets to train classifiers, gpt-oss-safeguard allows for real-time policy adjustments, making it easier for developers to refine their safety measures. This flexibility is crucial for applications where the nature of potential harm is constantly changing, such as in gaming or online product reviews. The model takes two inputs: a policy and the content to evaluate. It produces a classification and provides reasoning behind its conclusions. This approach works well in nuanced domains or when developers lack enough data to train effective classifiers. The flexibility in using any policy allows for more tailored safety measures, addressing the unique challenges of various platforms. OpenAI aims to gather feedback from the research and safety communities with the release of this model. Collaborating with ROOST, they plan to test its performance and document findings. The gpt-oss-safeguard represents a shift from traditional safety classifiers that require extensive manual curation of training data. Instead, it gives developers the tools to dynamically adapt safety measures, significantly improving response times and relevance to specific use cases.