Jailbreak evaluation on Synthetic dataset (held-out)
100Good Bot RateAIM
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| AIMModel=Claude v1.32023.07 | 100 | — | 0 | |
| AIMModel=GPT-42023.07 | 13 | — | 1 | |
| combination_3Model=Claude v1.32023.07 | 12 | — | 2 | |
| Adaptive attackModel=GPT-42023.07 | 4 | 96 | — | |
| combination_2Model=GPT-42023.07 | 3 | — | 10 | |
| combination_2Model=Claude v1.32023.07 | 3 | — | 8 | |
| combination_3Model=GPT-42023.07 | 2 | — | 5 | |
| Adaptive attackModel=Claude v1.32023.07 | 1 | 99 | — |