Safety IndexOpenAIgpt-oss-20b
gpt-oss-20b
OpenAIOpen weightsReleased 2025-08Reasoning model
Summary
gpt-oss-20b has data in 2 of the five pillars, short of the 4 needed for a headline number, so it is listed as partly measured; the mean of its measured pillars is 85. Its strongest area is interaction and honesty at 87, and its weakest is refusal calibration at 84.
3 measurements count toward the score, 0 of them from SAI's own runs. Not yet measured: adversarial robustness, agentic safety, people and society.
Pillar scores
How gpt-oss-20b compares with every other model scored in the pillar, and which sub-categories the score rests on.
Refusal calibration
Does the model refuse what it should and help with what it may?
- Harmful requests84 from 2 evaluations
- Over-refusalnot measured
Rank 15 of 39 scored models.
Adversarial robustness
Do its refusals and instructions hold up when someone tries to get around them?
- Jailbreaksnot measured
- Prompt injectionnot measured
- Multilingual safetynot measured
9 models scored.
Agentic safety
Does it behave safely when it has tools and autonomy?
- Agent misusenot measured
- Reward hacking and task integritynot measured
- Overreach and destructive actionsnot measured
3 models scored.
Interaction and honesty
Is it honest with the person it is talking to?
- Factuality and hallucinationnot measured
- Honesty under pressure and sycophancy87 from 1 evaluation
- Deception and schemingnot measured
- Alignment auditnot measured
- Manipulation of the usernot measured
Rank 6 of 29 scored models.
People and society
Does it protect vulnerable people and the public interest?
- Self-harm and wellbeingnot measured
- Child safetynot measured
- Bias, fairness and neutralitynot measured
- Privacy: memorisation and leakagenot measured
- Privacy: surveillance assistancenot measured
- Elections and influence operationsnot measured
3 models scored.
Every measurement
These are the raw numbers behind the scores, with who measured them, when, and how. Rows marked counted feed the index. The others are shown for context, such as benign twins of harm rows, retired evaluations that every frontier model passes, tests of dangerous knowledge, other evaluators' overall scores, and the developer's own figures.
| Evaluation | Value | Score | Source | Setting |
|---|---|---|---|---|
| Refusal calibration | ||||
| Harmful requests | ||||
| AIR-Bench 2024 | 86.0%n = 5,694 | 86 | Stanford CRFM HELM AIR-Bench v1.19.0 Independent evaluator | Published by Stanford CRFM HELM AIR-Bench v1.19.0.Full configurationfull 5,694-prompt set, Stanford HELM AIR-Bench v1.19.0; overall refusal rate, 5,694 prompts, GPT-4o judge; HELM row 'gpt-oss-20b' |
| Scale FORTRESS (risk score) | 17.6%95% CI 14.8% to 20.4% | 82 | Scale AI Fortress leaderboard Independent evaluator | Published by Scale AI Fortress leaderboard.Full configurationFortress row 'gpt-oss-20b' |
| HarmBench (direct requests) | 98.7% | retired, not counted | Stanford CRFM HELM Safety v1.17.0 Independent evaluator · 2025-11-24 | Published by Stanford CRFM HELM Safety v1.17.0 (2025-11-24).Full configurationHELM Safety score 0-100, higher is safer (not an attack-success rate); HELM Safety v1.17.0; HELM row 'gpt-oss-20b' |
| Over-refusal | ||||
| XSTest (HELM Safety) | 87.2% | retired, not counted | Stanford CRFM HELM Safety v1.17.0 Independent evaluator · 2025-11-24 | Published by Stanford CRFM HELM Safety v1.17.0 (2025-11-24).Full configurationHELM Safety v1.17.0; HELM row 'gpt-oss-20b' |
| Interaction and honesty | ||||
| Honesty under pressure and sycophancy | ||||
| MASK | 86.5%95% CI 84.4% to 88.5%n = 500 | 87 | Scale AI SEAL / CAIS MASK leaderboard Independent evaluator | Published by Scale AI SEAL / CAIS MASK leaderboard.Full configurationScale leaderboard on the private 500-example split, not the public split; MASK private split, 500 examples; honesty score = 1 - P(lie); row 'gpt-oss-20b' |
| People and society | ||||
| Bias, fairness and neutrality | ||||
| BBQ | 96.7%n = 1,000 | retired, not counted | Stanford CRFM HELM Safety v1.17.0 Independent evaluator · 2025-11-24 | Published by Stanford CRFM HELM Safety v1.17.0 (2025-11-24).Full configurationHELM Safety v1.17.0; HELM row 'gpt-oss-20b' |
| Other evaluators' scores (not counted) | ||||
| Anthropic Red Team prompts (HELM Safety) | 99.7% | context, not counted | Stanford CRFM HELM Safety v1.17.0 Independent evaluator · 2025-11-24 | Published by Stanford CRFM HELM Safety v1.17.0 (2025-11-24).Full configurationHELM Safety v1.17.0; HELM row 'gpt-oss-20b' |
| HELM Safety mean score | 96.4% | context, not counted | Stanford CRFM HELM Safety v1.17.0 Independent evaluator · 2025-11-24 | Published by Stanford CRFM HELM Safety v1.17.0 (2025-11-24).Full configurationHELM Safety v1.17.0; HELM row 'gpt-oss-20b' |
| SimpleSafetyTests (HELM Safety) | 100% | context, not counted | Stanford CRFM HELM Safety v1.17.0 Independent evaluator · 2025-11-24 | Published by Stanford CRFM HELM Safety v1.17.0 (2025-11-24).Full configurationHELM Safety v1.17.0; HELM row 'gpt-oss-20b' |
| Developer-reported (not counted) | ||||
| BBQ ambiguous accuracy (lab-reported) | 79% | developer-reported, not counted | OpenAI (gpt-oss model card) Developer-reported · 2025-08-05 | Reported by the model's developer in its own system card or report.Full configurationBBQ ambiguous accuracy; Table 10 |