Invocation attack on ToolBench
98CDAFlip-Agent
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Flip-AgentModel=Llama-3.2-3B-Instruct2026.03 | 98 | 98.9 | |
| Flip-AgentModel=ToolLLaMA-2-7b-v22026.03 | 98 | 100 | |
| Flip-SModel=ToolLLaMA-2-7b-v22026.03 | 97 | 100 | |
| Flip-AgentModel=Qwen3-8B2026.03 | 97 | 67.3 | |
| Flip-SModel=Qwen3-4B-Instruct-25072026.03 | 96 | 76.8 | |
| TBTModel=ToolLLaMA-2-7b-v22026.03 | 95 | 0 | |
| TrojViTModel=ToolLLaMA-2-7b-v22026.03 | 95 | 40 | |
| Flip-AgentModel=Llama-3.1-8B-Instruct2026.03 | 95 | 97.2 | |
| TBTModel=Qwen3-8B2026.03 | 95 | 0 | |
| Flip-AgentModel=Qwen3-4B-Instruct-25072026.03 | 95 | 87.3 | |
| TrojViTModel=Llama-3.1-8B-Instruct2026.03 | 92 | 65 | |
| TrojViTModel=Llama-3.2-3B-Instruct2026.03 | 90 | 55 | |
| Flip-SModel=Llama-3.1-8B-Instruct2026.03 | 90 | 96.8 | |
| TrojViTModel=Qwen3-8B2026.03 | 90 | 60 | |
| TBTModel=DeepSeek-R1-Distill-Llama-8B2026.03 | 90 | 5 | |
| TrojViTModel=DeepSeek-R1-Distill-Llama-8B2026.03 | 90 | 60 | |
| Flip-AgentModel=DeepSeek-R1-Distill-Llama-8B2026.03 | 90 | 82.1 | |
| Flip-SModel=Llama-3.2-3B-Instruct2026.03 | 85 | 83.1 | |
| TrojViTModel=Qwen3-4B-Instruct-25072026.03 | 85 | 60 | |
| Flip-SModel=DeepSeek-R1-Distill-Llama-8B2026.03 | 85 | 76.8 | |
| Flip-SModel=Qwen3-8B2026.03 | 81 | 63.1 | |
| TBTModel=Llama-3.2-3B-Instruct2026.03 | 78 | 0 | |
| TBTModel=Llama-3.1-8B-Instruct2026.03 | 70 | 25 | |
| TBTModel=Qwen3-4B-Instruct-25072026.03 | 70 | 10 |