Safety Evaluation on HarmBench Contextual Trajectory Evaluation Multi-turn
96.1SFRSTAR
Evaluation Results
| Method | Links | |
|---|---|---|
| STARTarget Model=Gemini 2.0-Flash, Access Type=Closed-Source Models2026.03 | 96.1 | |
| STARTarget Model=GPT 4o, Access Type=Closed-Source Models2026.03 | 94.5 | |
| X-teamingTarget Model=GPT 4o, Access Type=Closed-Source Models2026.03 | 94.3 | |
| STARTarget Model=Llama 3-8B-IT, Access Type=Open-Weight Models2026.03 | 89 | |
| X-teamingTarget Model=Gemini 2.0-Flash, Access Type=Closed-Source Models2026.03 | 87.4 | |
| X-teamingTarget Model=Llama 3-8B-IT, Access Type=Open-Weight Models2026.03 | 85.5 | |
| STARTarget Model=Llama 3-70B-IT, Access Type=Open-Weight Models2026.03 | 85.5 | |
| ActorAttackTarget Model=Llama 3-70B-IT, Access Type=Open-Weight Models2026.03 | 85 | |
| X-teamingTarget Model=Llama 3-70B-IT, Access Type=Open-Weight Models2026.03 | 84.9 | |
| ActorAttackTarget Model=GPT 4o, Access Type=Closed-Source Models2026.03 | 84.5 | |
| RACETarget Model=GPT 4o, Access Type=Closed-Source Models2026.03 | 82.8 | |
| ActorAttackTarget Model=Llama 3-8B-IT, Access Type=Open-Weight Models2026.03 | 79 | |
| STARTarget Model=Claude 3.5 Sonnet, Access Type=Closed-Source Models2026.03 | 74 | |
| X-teamingTarget Model=Claude 3.5 Sonnet, Access Type=Closed-Source Models2026.03 | 67.9 | |
| ActorAttackTarget Model=Claude 3.5 Sonnet, Access Type=Closed-Source Models2026.03 | 66.5 | |
| CrescendoTarget Model=Llama 3-70B-IT, Access Type=Open-Weight Models2026.03 | 62 | |
| CrescendoTarget Model=Llama 3-8B-IT, Access Type=Open-Weight Models2026.03 | 60 | |
| CrescendoTarget Model=Claude 3.5 Sonnet, Access Type=Closed-Source Models2026.03 | 50 | |
| CrescendoTarget Model=GPT 4o, Access Type=Closed-Source Models2026.03 | 46 | |
| ActorAttackTarget Model=Gemini 2.0-Flash, Access Type=Closed-Source Models2026.03 | 42.1 | |
| CoATarget Model=Llama 3-8B-IT, Access Type=Open-Weight Models2026.03 | 25.5 | |
| CoATarget Model=Llama 3-70B-IT, Access Type=Open-Weight Models2026.03 | 18.8 | |
| CoATarget Model=GPT 4o, Access Type=Closed-Source Models2026.03 | 17.5 | |
| CoATarget Model=Claude 3.5 Sonnet, Access Type=Closed-Source Models2026.03 | 3.4 |