Prompt Injection Attack on MuSiQue (ASR, Memory, Time)
98Attack Success Rate (ASR)nanoGCG-OPT
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| nanoGCG-OPTTarget Model=Meta-SecAlign-8B2026.04 | 98 | 82.7 | 1,652.4 | |
| FlashRTTarget Model=Meta-SecAlign-8B2026.04 | 98 | 35.8 | 782.2 | |
| nanoGCG-OPTLLM=Llama-3.1-8B-Instruct2026.04 | 94 | 91.9 | 4,209.8 | |
| FlashRTLLM=Llama-3.1-8B-Instruct2026.04 | 94 | 41.6 | 959 | |
| nanoGCGTarget Model=Meta-SecAlign-8B2026.04 | 86 | 81.8 | 1,810.5 | |
| nanoGCGLLM=Llama-3.1-8B-Instruct2026.04 | 80 | 89.7 | 3,201.2 | |
| Context ClippingLLM=Llama-3.1-8B-Instruct2026.04 | 54 | 41.1 | 250.8 | |
| Heuristic AttackLLM=Llama-3.1-8B-Instruct2026.04 | 32 | 0 | 0 | |
| HeuristicTarget Model=Meta-SecAlign-8B2026.04 | 14 | 0 | 0 |