OpenAI gpt-oss-safeguard Technical Report

OpenAI has released gpt-oss-safeguard-120b and gpt-oss-safeguard-20b, open-weight reasoning models designed to classify content according to a provided policy. These models are post-trained from the gpt-oss family and are available under the Apache 2.0 license.

Model Versions and Licensing

gpt-oss-safeguard models are available in two sizes: 120b and 20b. These models are text-only and are compatible with the Responses API. They are released under the Apache 2.0 license and OpenAI's gpt-oss usage policy.

Technical Capabilities and Customization

These models are designed for policy-based content classification. They are optimized to reason through a provided policy to determine if content should be labeled accordingly. Key technical features include:

  • Chain-of-Thought (CoT): The models provide full chain-of-thought reasoning, allowing users to see the internal logic used to reach a classification decision.

  • Reasoning Effort Levels: Users can configure the models with different reasoning efforts (low, medium, high) to balance performance and latency.

  • Structured Outputs: The models support Structured Outputs, ensuring that classification labels are returned in a predictable, machine-readable format.

Intended Use and Application

OpenAI recommends using gpt-oss-safeguard models as a classification layer rather than as the core user-facing functionality of an application. For general-purpose interaction, the original gpt-oss models are better suited.

Safety and Risk Assessment

Because these models are open-weight, OpenAI conducted safety evaluations to ensure they meet safety standards even if used in a chat setting, despite this not being their intended primary use case.

Regarding frontier risks, OpenAI states that the models were fine-tuned from gpt-oss without any additional biological or cybersecurity data. Consequently, the risk assessments performed during the original gpt-oss release—specifically the work on estimating worst-case scenarios—cross-apply to the gpt-oss-safeguard series.

Sources