Adversarial Attack on 1,000-pair main panel
92.7ASRFRA-Attack
Evaluation results
| Method | Links | ||
|---|---|---|---|
| FRA-AttackTarget Model=Claude-Opus-4.6-thinking, GPTScore success threshold=0.32026.05 | 92.7 | 61.4 | |
| M-Attack-V2Target Model=Claude-Opus-4.6-thinking, GPTScore success threshold=0.32026.05 | 89.3 | 54.7 | |
| FOA-AttackTarget Model=Claude-Opus-4.6-thinking, GPTScore success threshold=0.32026.05 | 82.6 | 45.3 | |
| M-Attack-V2Target Model=Gemini-3-flash-thinking, GPTScore success threshold=0.32026.05 | 80.5 | 42.6 | |
| FRA-AttackTarget Model=Gemini-3-flash-thinking, GPTScore success threshold=0.32026.05 | 79.7 | 45.6 | |
| M-AttackTarget Model=Claude-Opus-4.6-thinking, GPTScore success threshold=0.32026.05 | 77.1 | 41.4 | |
| FRA-AttackTarget Model=GPT-5.4-thinking, GPTScore success threshold=0.32026.05 | 72.9 | 41.6 | |
| M-Attack-V2Target Model=GPT-5.4-thinking, GPTScore success threshold=0.32026.05 | 70 | 37.7 | |
| FOA-AttackTarget Model=Gemini-3-flash-thinking, GPTScore success threshold=0.32026.05 | 68.1 | 34.8 | |
| M-AttackTarget Model=Gemini-3-flash-thinking, GPTScore success threshold=0.32026.05 | 61.4 | 30.6 | |
| FOA-AttackTarget Model=GPT-5.4-thinking, GPTScore success threshold=0.32026.05 | 58.5 | 30.1 | |
| M-AttackTarget Model=GPT-5.4-thinking, GPTScore success threshold=0.32026.05 | 50.9 | 26.4 |