SSP-Bench - Safety: leaderboard
The safety service of SSP-Bench (Deniz, Boshmaf and Khalil, QCRI, 2026), a dynamic benchmark that generates items from external sources and selects them for difficulty and separability. Each model answers 1,768 generated harmful prompts grounded in AIAAIC incident reports and red-teaming corpora, including mutated variants, and a gpt-oss-safeguard-120b judge labels each response safe or unsafe. Safety rate in percent; higher is better.
Metric: Safety rate (%; share of the 1,768 generated harmful prompts, AIAAIC-grounded and mutated red-teaming items, that the model refuses or answers safely, judged by gpt-oss-safeguard-120b; SSP-Bench dynamic benchmark instance scored on the 24-model testing panel, held out from item generation and selection; proprietary models through OpenRouter, March-April 2026). Source: arxiv.org. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-OSS-120B | 99.4 | #699 |
| 2 | Claude 3.5 Haiku | 99.3 | #1563 |
| 3 | Qwen 3.5 27B | 99.1 | #358 |
| 4 | Qwen 3.5 122B A10B | 98.9 | #329 |
| 5 | Qwen 3.5 9B | 98.7 | #803 |
| 6 | Qwen 3.5 4B | 98.6 | #1180 |
| 7 | Qwen 3.5 35B A3B | 98.6 | #480 |
| 8 | GPT-OSS-20B | 97.1 | #1150 |
| 9 | GPT-5 Mini | 96.9 | #349 |
| 10 | Gemma 3 1B | 93.8 | #9004 |
| 11 | Claude Haiku 4.5 | 93.2 | #568 |
| 12 | Gemma 3 4B | 91.4 | #5104 |
| 13 | Phi-4 | 86.7 | #1920 |
| 14 | Gemma 3 27B | 85.5 | #1638 |
| 15 | Gemma 3 12B | 84.4 | #2013 |
Interactive version: theaggregate.ai/benchmark?slug=ssp-bench-safety · How It Works · Data refreshed daily, snapshot 2026-11-12.