AdvBench: leaderboard
Adversarial behavior dataset widely used for evaluating jailbreak attacks and safety bypasses, especially in work on adversarial suffixes and harmful behavior elicitation.
Metric: AdvBench harmful-instruction score. Source: github.com. Source: benchmarklist.com. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 2 7B Chat | 1 | #1499 |
| 2 | Llama 2 13B Chat Base | 1 | #1422 |
| 3 | Llama 2 70B Chat (HF) | 1 | #1264 |
| 4 | Qwen 1.5 72B Chat | 1 | #776 |
| 5 | GPT-4 Preview (1106) | 1 | #418 |
| 6 | Mistral 7B Instruct (v0.3) | 1 | #1274 |
| 7 | Claude 3 Opus | 0.99 | #571 |
| 8 | Mixtral 8x7B Instruct (v0.1) | 0.99 | #1073 |
| 9 | Yi Chat 34B | 0.99 | #976 |
| 10 | Mistral 7B Instruct (v0.2) | 0.99 | #1333 |
| 11 | Gemini 1.5 Flash (001) | 0.99 | #577 |
| 12 | GPT-3.5 Turbo (0125) | 0.99 | #984 |
| 13 | Bielik-11B-v2.3-Instruct | 0.99 | #693 |
| 14 | Mistral-7B-v0.3 | 0.98 | #1355 |
| 15 | Gemma 2 9B | 0.93 | #958 |
Interactive version: theaggregate.ai/benchmark?slug=advbench · How It Works · Data refreshed daily, snapshot 2026-10-20.