SALAD-Bench Attack: leaderboard
SALAD-Bench — hierarchical LLM safety benchmark; attack-enhanced (jailbreak-augmented) question set scored by MD-Judge safety rate.
Metric: Safety Score (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 34 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 2 | 88.06 | #761 |
| 2 | GPT-4 Turbo | 80.28 | #404 |
| 3 | GPT-3.5 Turbo | 73.4 | #865 |
| 4 | Llama 2 70B Chat (HF) | 66.24 | #1264 |
| 5 | Llama 2 13B Chat Base | 65.72 | #1422 |
| 6 | gemma-2B (IT) | 50.34 | #1559 |
| 7 | Yi Chat 34B | 23.64 | #976 |
| 8 | Qwen 1.5 4B Chat | 23.34 | #1467 |
| 9 | Yi 6B (Chat) | 22.56 | #1408 |
| 10 | internlm2-chat-7B | 20.28 | #1121 |
| 11 | Gemini 1.0 Pro | 20.04 | #933 |
| 12 | Llama 2 7B Chat | 18.24 | #1499 |
| 13 | Qwen 1.5 72B Chat | 16.22 | #776 |
| 14 | gemma-7B (IT) | 15.54 | #1496 |
| 15 | chatglm3-6B | 12.48 | #1439 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=salad-bench-attack · How It Works · Data refreshed daily, snapshot 2026-10-20.