Safety evaluation against dynamic adversarial chains on DREAM (test)
67.3Overall Defense ScoreClaude-Sonnet-4.5
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Claude-Sonnet-4.5Source=Proprietary (Closed-Source)2025.12 | 67.3 | 24 | 76.8 | 71.95 | 84.23 | |
| Grok-4Source=Proprietary (Closed-Source)2025.12 | 61.17 | 43 | 75.32 | 64.68 | 67.53 | |
| DeepSeek-V3.1Source=Open-Source Models2025.12 | 48.1 | 67 | 65.49 | 54.39 | 49.48 | |
| Gemini-2.5-ProSource=Proprietary (Closed-Source)2025.12 | 45.99 | 72 | 64.99 | 51.49 | 49.75 | |
| GPT-5Source=Proprietary (Closed-Source)2025.12 | 45.39 | 62 | 62.85 | 52.39 | 52 | |
| Qwen-2.5-72BSource=Open-Source Models2025.12 | 43.37 | 73 | 61.61 | 48.65 | 44.99 | |
| Llama-3.1-70BSource=Open-Source Models2025.12 | 41.4 | 76 | 58.43 | 48.21 | 43.91 | |
| Gemini-2.5-FlashSource=Proprietary (Closed-Source)2025.12 | 40.89 | 76 | 60.27 | 47.41 | 44.9 | |
| Kimi-K2-PreviewSource=Open-Source Models2025.12 | 39.5 | 77 | 59.61 | 47.07 | 46.4 | |
| o4-miniSource=Proprietary (Closed-Source)2025.12 | 39.04 | 75 | 57.05 | 47.49 | 46.53 | |
| Qwen3-235B-A22BSource=Open-Source Models2025.12 | 38.73 | 83 | 58.1 | 44.97 | 37.82 | |
| Gemini-2.5-Flash-NTSource=Proprietary (Closed-Source)2025.12 | 37.34 | 82 | 56.32 | 44.8 | 41.37 |