Agentic Attack on WhisperBench Background execution (held-out)
96.4ISRMEMGHOST
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| MEMGHOSTTarget Model=GPT-5.4, Evaluation Cases=562026.07 | 96.4 | 100 | 87.5 | 87.5 | |
| MEMGHOSTTarget Model=Grok 4.1 Fast, Evaluation Cases=562026.07 | 96.4 | 100 | 91.1 | 91.1 | |
| MEMGHOSTTarget Model=DeepSeek v4 Flash, Evaluation Cases=562026.07 | 96.4 | 87.5 | 85.7 | 75 | |
| MEMGHOSTTarget Model=MiniMax-M2.5, Evaluation Cases=562026.07 | 94.6 | 94.6 | 82.1 | 78.6 | |
| MEMGHOSTTarget Model=Sonnet 4.5, Evaluation Cases=562026.07 | 92.9 | 94.6 | 82.1 | 80.4 | |
| MEMGHOSTTarget Model=Sonnet 4.6, Evaluation Cases=562026.07 | 73.2 | 73.2 | 71.4 | 71.4 | |
| RGSTarget Model=GPT-5.4, Evaluation Cases=202026.07 | 40 | 50 | 40 | 40 | |
| MAP-EliteTarget Model=GPT-5.4, Evaluation Cases=202026.07 | 20 | 5 | 20 | 0 | |
| Naive AttackTarget Model=GPT-5.4, Evaluation Cases=562026.07 | 0 | 0 | 0 | 0 | |
| Escape CharactersTarget Model=GPT-5.4, Evaluation Cases=562026.07 | 0 | 5.4 | 0 | 0 | |
| Context IgnoreTarget Model=GPT-5.4, Evaluation Cases=562026.07 | 0 | 25 | 0 | 0 | |
| Fake CompletionTarget Model=GPT-5.4, Evaluation Cases=562026.07 | 0 | 5.4 | 0 | 0 | |
| Combined AttackTarget Model=GPT-5.4, Evaluation Cases=562026.07 | 0 | 19.6 | 0 | 0 |