Indirect Prompt Injection on HotpotQA
100ASRObliInjection-CE
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ObliInjection-CETarget LLM=Mistral-7B2025.12 | 100 | — | |
| ObliInjection-CETarget LLM=Qwen-2.5-7B2025.12 | 100 | — | |
| ObliInjectionTarget LLM=Llama-3-8B2025.12 | 100 | — | |
| ObliInjectionTarget LLM=Mistral-7B2025.12 | 100 | — | |
| ObliInjectionTarget LLM=Falcon3-7B2025.12 | 100 | — | |
| ObliInjectionTarget LLM=Qwen3-4B2025.12 | 100 | — | |
| ObliInjection-CETarget LLM=Llama4-17B2025.12 | 99.8 | — | |
| ObliInjectionTarget LLM=Llama4-17B2025.12 | 99.6 | — | |
| ObliInjectionTarget LLM=Llama-3.1-8B2025.12 | 99.2 | — | |
| ObliInjectionTarget LLM=Qwen-2.5-7B2025.12 | 98.6 | — | |
| JudgeDeceiverTarget LLM=Qwen-2.5-7B2025.12 | 93.4 | — | |
| ObliInjection-CETarget LLM=Llama-3.1-8B2025.12 | 93 | — | |
| ObliInjection-GCGTarget LLM=Mistral-7B2025.12 | 92.6 | — | |
| ObliInjection-CETarget LLM=Falcon3-7B2025.12 | 91.6 | — | |
| ObliInjection-GCGTarget LLM=Falcon3-7B2025.12 | 88.6 | — | |
| JudgeDeceiverTarget LLM=Mistral-7B2025.12 | 88 | — | |
| ObliInjection-GCGTarget LLM=Llama-3-8B2025.12 | 85.4 | — | |
| ObliInjection-CETarget LLM=Qwen3-4B2025.12 | 84.6 | — | |
| Combined AttackTarget LLM=Qwen-2.5-7B2025.12 | 84.4 | — | |
| ObliInjection-GCGTarget LLM=Qwen-2.5-7B2025.12 | 70 | — | |
| ObliInjection-CETarget LLM=Llama-3-8B2025.12 | 67.8 | — | |
| JudgeDeceiverTarget LLM=Falcon3-7B2025.12 | 58.8 | — | |
| JudgeDeceiverTarget LLM=Llama-3-8B2025.12 | 55.6 | — | |
| Combined AttackTarget LLM=Llama-3.1-8B2025.12 | 39.6 | — | |
| Neural ExecTarget LLM=Llama-3-8B2025.12 | 32.8 | — | |
| JudgeDeceiverTarget LLM=Llama-3.1-8B2025.12 | 32.8 | — | |
| ObliInjection-GCGTarget LLM=Llama-3.1-8B2025.12 | 31.2 | — | |
| Combined AttackTarget LLM=Qwen3-4B2025.12 | 28.4 | — | |
| Neural ExecTarget LLM=Qwen3-4B2025.12 | 27.8 | — | |
| Neural ExecTarget LLM=Mistral-7B2025.12 | 25.5 | — | |
| Neural ExecTarget LLM=Llama-3.1-8B2025.12 | 25.4 | — | |
| Neural ExecTarget LLM=Falcon3-7B2025.12 | 20.2 | — | |
| Combined AttackTarget LLM=Mistral-7B2025.12 | 20 | — | |
| Combined AttackTarget LLM=Llama-3-8B2025.12 | 14.2 | — | |
| ObliInjection-GCGTarget LLM=Qwen3-4B2025.12 | 14 | — | |
| JudgeDeceiverTarget LLM=Qwen3-4B2025.12 | 10.6 | — | |
| Neural ExecTarget LLM=Llama4-17B2025.12 | 8.8 | — | |
| ObliInjection-GCGTarget LLM=Llama4-17B2025.12 | 8.2 | — | |
| Neural ExecTarget LLM=Qwen-2.5-7B2025.12 | 2.8 | — | |
| Combined AttackTarget LLM=Llama4-17B2025.12 | 1.6 | — | |
| Combined AttackTarget LLM=Falcon3-7B2025.12 | 0.4 | — | |
| JudgeDeceiverTarget LLM=Llama4-17B2025.12 | 0.2 | — | |
| Combined AttackTarget LLM=Average2025.12 | — | 26.9 | |
| JudgeDeceiverTarget LLM=Average2025.12 | — | 48.5 | |
| Neural ExecTarget LLM=Average2025.12 | — | 20.5 | |
| ObliInjectionTarget LLM=Average2025.12 | — | 99.6 | |
| ObliInjection-CETarget LLM=Average2025.12 | — | 91 | |
| ObliInjection-GCGTarget LLM=Average2025.12 | — | 55.7 |