Agent-SafetyBench - Content: leaderboard
The cases without environments, where unsafe means harmful text output; the number is the share judged safe.
Metric: Safety score (% safe, higher is better). Source: arxiv.org. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.5 Sonnet | 88.6 | #746 |
| 2 | Claude 3.5 Haiku | 86.4 | #1553 |
| 3 | Claude 3 Opus | 84.9 | #1455 |
| 4 | Llama 3.1 405B Instruct | 79.6 | #1115 |
| 5 | GPT-4 Turbo | 72.7 | #1001 |
| 6 | GPT-4o | 72.5 | #741 |
| 7 | GPT-4o Mini | 72.5 | #1623 |
| 8 | Qwen 2.5 72B Instruct | 71 | #1092 |
| 9 | Llama 3.1 70B Instruct | 69.8 | #1546 |
| 10 | Gemini 1.5 Pro | 69.3 | #1049 |
| 11 | Gemini 1.5 Flash | 69.1 | #1668 |
| 12 | Qwen 2.5 14B Instruct | 60.6 | #1732 |
| 13 | Llama 3.1 8B Instruct | 58.6 | #4246 |
| 14 | DeepSeek V2.5 | 55.7 | #1210 |
| 15 | GLM-4 9B Chat | 44.3 | #2759 |
Interactive version: theaggregate.ai/benchmark?slug=agent-safetybench-content · How It Works · Data refreshed daily, snapshot 2026-11-13.