| Model,Overall Safe Tasks Score,Overall Harm Tasks Score,Human-Generated Safe Tasks Score,Human-Generated Harm Tasks Score,HITL Safe Tasks Score,HITL Harm Tasks Score,Refusal Rate,Normalized Safety Score,Open,Bias,Cybercrime,Harassment,Misinformation,Illegal Activity | |
| GPT-4o,34.4,22.8,44.0,26.0,28.0,20.7,30.2,31.7,False,14.0,16.0,16.0,28.0,40.0 | |
| GPT-4o-Mini,18.4,14.0,22.0,14.0,16.0,14.0,36.5,35.7,False,6.0,8.0,14.0,24.0,18.0 | |
| Claude-3.5-Sonnet,21.2,7.6,24.0,12.0,19.3,4.7,57.7,55.0,False,4.0,6.0,5.0,12.0,12.0 | |
| Llama-3.2-90B,8.4,11.2,10.1,10.0,7.3,12.0,14.0,34.0,True,22.0,8.0,10.0,14.0,2.0 | |
| Qwen-2-VL-72B,24.4,26.0,30.0,29.0,20.7,24.0,0.8,21.5,True,34.0,18.0,18.0,30.0,30.0 |