Skip to main content
Lab publication Credible

gpt-oss-safeguard: Open-Weight Safety Reasoning Models

OpenAI released gpt-oss-safeguard, a pair of open-weight safety-classification models (120B and 20B) under an Apache 2.0 license. The models take a safety policy supplied by the developer at inference time and classify content against it, emitting chain-of-thought reasoning alongside the classification rather than a fixed-taxonomy score. They are presented as an open-weight counterpart to an internal safety-reasoning system.

Publisher

OpenAI

Published

29 Oct 2025

Added

today

DOI

Key Findings

  • Policy-conditional design: the safety taxonomy is supplied by the developer at inference time rather than fixed in the model
  • Outputs explicit reasoning alongside each safety classification
  • Released as open weights (Apache 2.0) in 120B and 20B sizes; developed with partners including Discord, SafetyKit, and ROOST

Methodology Notes

Model card / research-preview release, 29 October 2025. The accompanying OpenAI announcement and technical report were published the same period but return access errors to automated fetchers, so the Hugging Face model card is the verifiable primary source. Note: an August 2025 arXiv report (2508.10925) covers the base gpt-oss models, not this safeguard variant.

Tags

gpt-oss-safeguardopenaiguardrail-modelpolicy-conditionalopen-weightsroost

Cite This

APA

OpenAI (2025). gpt-oss-safeguard: Open-Weight Safety Reasoning Models. OpenAI. https://huggingface.co/openai/gpt-oss-safeguard-120b