Agent-SafetyBench - Behavior: leaderboard
The cases with tool environments, where unsafe means an unsafe tool action; 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 Opus | 53.2 | #1455 |
| 2 | Claude 3.5 Sonnet | 51.9 | #746 |
| 3 | Claude 3.5 Haiku | 40.7 | #1553 |
| 4 | GPT-4o | 36.9 | #741 |
| 5 | Gemini 1.5 Flash | 34.6 | #1668 |
| 6 | GLM-4 9B Chat | 34.6 | #2759 |
| 7 | GPT-4 Turbo | 33.9 | #1001 |
| 8 | Gemini 1.5 Pro | 29.2 | #1049 |
| 9 | Qwen 2.5 72B Instruct | 28.6 | #1092 |
| 10 | DeepSeek V2.5 | 28.6 | #1210 |
| 11 | Qwen 2.5 14B Instruct | 24.4 | #1732 |
| 12 | Llama 3.1 405B Instruct | 24 | #1115 |
| 13 | Llama 3.1 70B Instruct | 21.2 | #1546 |
| 14 | GPT-4o Mini | 20.5 | #1623 |
| 15 | Qwen 2.5 7B Instruct | 13.5 | #2564 |
Interactive version: theaggregate.ai/benchmark?slug=agent-safetybench-behavior · How It Works · Data refreshed daily, snapshot 2026-11-13.