In a significant move for AI safety and accessibility, OpenAI is collaborating with Ollama and ROOST (Robust Open Online Safety Tools) to release gpt-oss-safeguard, a suite of open-source models designed for safety classification tasks. This collaboration aims to empower developers and organizations with tools to better understand and mitigate potential harms associated with large language models. The release marks a pivotal step towards democratizing AI safety, offering a transparent and customizable approach to content moderation and policy enforcement. The open-source nature of gpt-oss-safeguard enables community-driven improvements and adaptation to diverse use cases, fostering a more responsible AI ecosystem.

The gpt-oss-safeguard models are specifically trained to reason about safety, making them suitable for applications like LLM input-output filtering, online content labeling, and offline analysis for trust and safety teams. Key highlights include:

The release of gpt-oss-safeguard is a game-changer for several reasons. Firstly, it addresses the critical need for accessible and customizable AI safety tools. Organizations can now leverage these models to enforce their own specific policies and adapt them to their unique contexts. Secondly, the open-source nature of the models fosters transparency and collaboration, enabling the community to contribute to their improvement and refinement. This is particularly important in the rapidly evolving landscape of AI safety, where continuous adaptation and innovation are essential. Finally, the ability to access the model's reasoning process provides valuable insights into its decision-making, increasing trust and enabling more effective debugging and refinement of safety policies. This ultimately empowers developers and safety teams to build safer and more responsible AI systems.

OpenAI evaluated the gpt-oss-safeguard models on both internal and external evaluation sets. The internal evaluation involved providing multiple policies simultaneously to the models and assessing their ability to correctly classify text under all included policies. This is a challenging task, as the model is only considered accurate if it exactly matches the golden set labels for all policies. OpenAI also evaluated the models on their moderation dataset released in 2022 and on ToxicChat, a public benchmark based on user queries to an open-source chatbot.

To run the models, users can download Ollama and execute the following commands in a terminal:

The models' performance is configurable, allowing users to adjust the reasoning effort based on their specific use case and latency requirements. This flexibility is crucial for adapting the models to diverse applications with varying performance constraints. The permissive Apache 2.0 license enables a wide range of use cases, from academic research to commercial deployment, without the restrictions associated with more restrictive licenses.

| Feature | Description | | --------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Model Sizes | 20B and 120B parameters | | Training Data | Trained and tuned for safety reasoning | | Policy Enforcement | Interprets user-defined policies | | Reasoning Access | Provides access to the model's reasoning process | | Reasoning Effort | Configurable (low, medium, high) | | License | Apache 2.0 | | Evaluation Datasets | OpenAI internal datasets, OpenAI moderation dataset (2022), ToxicChat |

The collaboration between OpenAI, Ollama, and ROOST to release gpt-oss-safeguard marks a significant step forward in democratizing AI safety. By providing open-source, customizable, and transparent safety models, this initiative empowers developers and organizations to build more responsible and trustworthy AI systems. As the field of AI safety continues to evolve, we anticipate further advancements in open-source tools and techniques that will contribute to a safer and more beneficial AI future.

Sources verified via Ollama of November 4, 2025.