Backdoor Attack Detection on AgentPoison LLM Agent
60PrecisionAttnTrace
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| AttnTraceTrigger optimization method=GCG2025.08 | 60 | 100 | 87 | 2 | |
| AttnTraceTrigger optimization method=BadChain2025.08 | 60 | 100 | 76 | 4 | |
| AttnTraceTrigger optimization method=AgentPoison2025.08 | 60 | 100 | 93 | 3 | |
| AttnTraceTrigger optimization method=AutoDAN2025.08 | 59 | 98 | 89 | 4 | |
| AttnTraceTrigger optimization method=CPA2025.08 | 56 | 94 | 86 | 7 |