Red-teaming on AdvBench (test)
88ASRAMIS
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| AMISAttacker Model=Llama-3.1-8B, Target Model=Claude-3.5-Haiku2025.11 | 88 | 42 | |
| ActorBreakerAttacker Model=GPT-4o-mini, Target Model=Claude-3.5-Haiku2025.11 | 48 | 23 | |
| Rainbow TeamingIterations=500, Target Model=Claude-3.5-Haiku2025.11 | 38 | 3 | |
| ActorBreakerAttacker Model=Llama-3.1-8B, Target Model=Claude-3.5-Haiku2025.11 | 36 | 18 | |
| Rainbow TeamingIterations=300, Target Model=Claude-3.5-Haiku2025.11 | 34 | 2 | |
| Rainbow TeamingIterations=100, Target Model=Claude-3.5-Haiku2025.11 | 10 | 0 | |
| ActorBreakerAttacker Model=Llama-3.1-70B, Target Model=Claude-3.5-Haiku2025.11 | 6 | 2 | |
| ActorBreakerAttacker Model=Claude-3.5-Haiku, Target Model=Claude-3.5-Haiku2025.11 | 4 | 4 |