AI Potluck
Product / UX / Safety & Guardrails

gpt-oss-safeguard

OpenAI

OpenAI's safety reasoning model, built on gpt-oss and released with Hugging Face and ROOST. Rather than emitting a fixed label it reasons over a safety policy supplied at call time and returns a reasoned decision, with configurable reasoning effort. It ships in 20b and 120b variants.

gpt-oss-safeguard (20b / 120b), released by OpenAI with Hugging Face and ROOST and distributed through the ROOST Model Community. Built on gpt-oss, it interprets a policy you supply rather than a fixed taxonomy and returns a reasoned decision. The card carries no training-data section at all. Verified 2026-08-13 via the HF model card.

Openness

3 medium confidence
3.0
weights
open(gpt-oss-safeguard-20b/120b on HF)
data
not-released
license
Apache-2.0(OSI)

Open weights under Apache-2.0 with no released training data and no published fine-tuning pipeline. The rung above requires both, so this scores 3, the same rung as every other model that ships permissive weights without a recipe. Notable as a policy-conditioned safety reasoning model rather than a fixed classifier, but that shapes capability, not openness.

Adoption

3 high confidence
3.0

100,632 downloads in the trailing 30 days for openai/gpt-oss-safeguard-20b, which bands at 100K-1M, level 3. That sits barely above the 100K boundary, so the level could move in either direction.

Capability

4 medium confidence
4.0

A different shape from fixed classifiers: reasons over an arbitrary supplied policy, which generalizes to novel categories at some latency cost.

  • https://huggingface.co/openai/gpt-oss-safeguard-20b recorded 2026-08-13

    Card describes gpt-oss-safeguard-120b and gpt-oss-safeguard-20b as 'safety reasoning models built-upon gpt-oss' that classify against a policy you supply, with a 'Bring your own policy' feature that 'interprets your written policy, so it generalizes across products and use cases', and reasoned rather than bare-label decisions.

Verified 2026-08-12