Prompt Injection Attack on AgentDojo 13 non-agent benchmarks
0Training QueriesTAP
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| TAPTarget Model=Meta-SecAlign-8B, Evaluation Setting=ASR@102026.03 | 0 | 82,110 | 82,110 | |
| PAIRTarget Model=Meta-SecAlign-8B, Evaluation Setting=ASR@102026.03 | 0 | 74,460 | 74,460 | |
| StrategyTarget Model=Meta-SecAlign-8B, Evaluation Setting=ASR@102026.03 | 0 | 194,140 | 194,140 | |
| Vanilla GRPOTarget Model=Meta-SecAlign-8B, Evaluation Setting=ASR@102026.03 | 15,000 | 17,000 | 32,000 | |
| PISmithTarget Model=Meta-SecAlign-8B, Evaluation Setting=ASR@102026.03 | 15,000 | 17,000 | 32,000 | |
| RL-HammerTarget Model=Meta-SecAlign-8B, Evaluation Setting=ASR@102026.03 | 30,000 | 17,000 | 47,000 |