Attack Detection on Extraction Attacks 105K sample set
100Detection RatePromptGuard 2
Evaluation Results
| Method | Links | |
|---|---|---|
| PromptGuard 2acronym=PG2026.02 | 100 | |
| LogReg (Ours)input=raw activations, evaluation_regime=LODO2026.02 | 79 | |
| Llama-as-Judgeacronym=LJ, prompting=zero-shot2026.02 | 31.8 | |
| LlamaGuardacronym=LG2026.02 | 15.2 |