Jailbreak Attack Evaluation on HarmBench (400 random samples)
0ASRLlama-2-7B-Chat (Original)
Evaluation results
| Method | Links | |
|---|---|---|
| Llama-2-7B-Chat (Original)Backbone=Llama-2-7B-Chat, Steering=None, Attack=None2026.02 | 0 | |
| Llama-3-8B-Instruct (Original)Backbone=Llama-3-8B-Instruct, Steering=None, Attack=None2026.02 | 2 | |
| Gemma-7B-it (Original)Backbone=Gemma-7B-it, Steering=None, Attack=None2026.02 | 3.25 | |
| STEER-JSON (Llama-2-7B-Chat)Backbone=Llama-2-7B-Chat, Steering=JSON, Attack=None2026.02 | 11.5 | |
| STEER-JSON (Gemma-7B-it)Backbone=Gemma-7B-it, Steering=JSON, Attack=None2026.02 | 15.5 | |
| STEER-COMPLIANCE (Llama-2-7B-Chat)Backbone=Llama-2-7B-Chat, Steering=Compliance, Attack=None2026.02 | 16 | |
| STEER-JSON (Llama-3-8B-Instruct)Backbone=Llama-3-8B-Instruct, Steering=JSON, Attack=None2026.02 | 20 | |
| STEER-COMPLIANCE (Gemma-7B-it)Backbone=Gemma-7B-it, Steering=Compliance, Attack=None2026.02 | 25.5 | |
| STEER-COMPLIANCE (Llama-3-8B-Instruct)Backbone=Llama-3-8B-Instruct, Steering=Compliance, Attack=None2026.02 | 38.5 | |
| CoP (Llama-3-8B-Instruct)Backbone=Llama-3-8B-Instruct, Steering=None, Attack=CoP2026.02 | 71 | |
| CoP (Gemma-7B-it)Backbone=Gemma-7B-it, Steering=None, Attack=CoP2026.02 | 71 | |
| CoP (Llama-2-7B-Chat)Backbone=Llama-2-7B-Chat, Steering=None, Attack=CoP2026.02 | 77 | |
| STEER-JSON + CoP (Llama-3-8B-Instruct)Backbone=Llama-3-8B-Instruct, Steering=JSON, Attack=CoP2026.02 | 79.5 | |
| STEER-COMPLIANCE + CoP (Llama-3-8B-Instruct)Backbone=Llama-3-8B-Instruct, Steering=Compliance, Attack=CoP2026.02 | 81.25 | |
| STEER-JSON + CoP (Llama-2-7B-Chat)Backbone=Llama-2-7B-Chat, Steering=JSON, Attack=CoP2026.02 | 88.75 | |
| STEER-JSON + CoP (Gemma-7B-it)Backbone=Gemma-7B-it, Steering=JSON, Attack=CoP2026.02 | 90.25 | |
| STEER-COMPLIANCE + CoP (Gemma-7B-it)Backbone=Gemma-7B-it, Steering=Compliance, Attack=CoP2026.02 | 93.5 | |
| STEER-COMPLIANCE + CoP (Llama-2-7B-Chat)Backbone=Llama-2-7B-Chat, Steering=Compliance, Attack=CoP2026.02 | 94.25 |