Jailbreak Attack on HEX-PHI
54.9ASRIT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| IT2026.02 | 54.9 | — | |
| RL2026.02 | 51.7 | — | |
| RL2026.02 | 38.3 | — | |
| IT2026.02 | 36.5 | — | |
| Circuit Breakers2026.02 | 13.4 | — | |
| Circuit Breakers2026.02 | 12.4 | — | |
| BSAFE+2026.02 | 5.7 | — | |
| BSAFE+2026.02 | 5 | — | |
| RLBF2026.02 | 4.3 | — | |
| RLBF2026.02 | 3.7 | — | |
| R2JVictim Model=GPT-3.5-turbo-0125, Avg. Training FLOPS=42012 GFLOPS2025.02 | 0.8818 | 4.15 | |
| R2JVictim Model=Llama-2-7b-chat, Avg. Training FLOPS=42012 GFLOPS2025.02 | 0.8273 | 3.97 | |
| RedEvalVictim Model=GPT-3.5-turbo-01252025.02 | 0.2303 | 2.06 | |
| InitialVictim Model=GPT-3.5-turbo-01252025.02 | 0.0455 | 1.12 | |
| InitialVictim Model=Llama-2-7b-chat2025.02 | 0.0273 | 1.06 | |
| RedEvalVictim Model=Llama-2-7b-chat2025.02 | 0.003 | 1.02 |