Attack Detection on Harmful Attacks 105K sample set
97.4Detection RateLlamaGuard
Evaluation Results
| Method | Links | |
|---|---|---|
| LlamaGuardacronym=LG2026.02 | 97.4 | |
| Llama-as-Judgeacronym=LJ, prompting=zero-shot2026.02 | 85.8 | |
| LogReg (Ours)input=raw activations, evaluation_regime=LODO2026.02 | 68 | |
| LogReg (Ours)input=raw activations, evaluation_regime=LODO2026.02 | 67 | |
| PromptGuard 2acronym=PG2026.02 | 37.3 | |
| PromptGuard 2acronym=PG2026.02 | 36.7 | |
| LlamaGuardacronym=LG2026.02 | 27.4 | |
| Llama-as-Judgeacronym=LJ, prompting=zero-shot2026.02 | 7.1 |