Kill Switch Evaluation on KILLBENCH Text-based Agent 1.0 (test)
84.73KSRIPI
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| IPIModel=Grok-4.3, neff=5212025.09 | 84.73 | 89.44 | 10 | 0 | 100 | 5.27 | |
| IPIModel=GPT-5.2, neff=4062025.09 | 72 | 97.54 | 1.82 | 0 | 91.95 | 26.18 | |
| AUTOGUARDModel=Grok-4.3, neff=5272025.09 | 71.27 | 74.38 | 22.89 | 1.65 | 94.33 | 4.18 | |
| AUTOGUARDModel=GPT-5.2, neff=4142025.09 | 66.73 | 88.65 | 8.18 | 0.36 | 79.18 | 24.73 | |
| WarningModel=GPT-5.2, neff=2842025.09 | 48.42 | 93.67 | 2.56 | 0.71 | 96.68 | 48.36 | |
| AUTOGUARDModel=Qwen3.6-35B-A3B, neff=5272025.09 | 45.82 | 47.82 | 41.98 | 8.02 | 59.96 | 4.18 | |
| IPIModel=Qwen3.6-35B-A3B, neff=5252025.09 | 42.85 | 44.88 | 41 | 11.64 | 75.36 | 4.51 | |
| WarningModel=Grok-4.3, neff=4062025.09 | 23.98 | 32.47 | 38.18 | 11.69 | 78.29 | 26.18 | |
| AUTOGUARDModel=Gemma4-31B-it, neff=5502025.09 | 22.36 | 22.36 | 73.65 | 3.98 | 81.65 | 0 | |
| IPIModel=Gemma4-31B-it, neff=5492025.09 | 20.4 | 20.44 | 69.44 | 9.98 | 88.73 | 0.18 | |
| WarningModel=Qwen3.6-35B-A3B, neff=5222025.09 | 12.2 | 12.85 | 80.02 | 2.73 | 52.76 | 5.09 | |
| GCGModel=Grok-4.3, neff=5452025.09 | 4.91 | 4.96 | 84.91 | 9.25 | 1.47 | 0.93 | |
| GCGModel=GPT-5.2, neff=3472025.09 | 4.91 | 7.78 | 57.64 | 0.55 | 0 | 36.91 | |
| WarningModel=Gemma4-31B-it, neff=5132025.09 | 4.55 | 4.87 | 84.02 | 4.71 | 30.62 | 6.73 | |
| GCGModel=Qwen3.6-35B-A3B, neff=5292025.09 | 1.82 | 1.89 | 90.89 | 3.44 | 0 | 3.85 | |
| GCGModel=Gemma4-31B-it, neff=5482025.09 | 0.56 | 0.57 | 97.29 | 1.82 | 0 | 0.33 |