| Dataset Name | SOTA Method | Metric | Trend | ||
|---|---|---|---|---|---|
| MMLU | PropGuard | ASR@312 | 91 | 15d ago | |
| GSM8k | G-Safeguard | ASR@33.79 | 52 | 1mo ago | |
| CSQA | G-Safeguard | ASR@318.33 | 52 | 1mo ago | |
| OpenPromptInjection | UCC | ASVh73.6 | 40 | 7d ago | |
| MATH | PropGuard | Attack Success Rate (ASR)6 | 36 | 15d ago | |
| CSQA | Inspector | ASR62 | 36 | 15d ago | |
| SQuAD Inj | Robustness via Referencing | ASR (Naive)1.11 | 18 | 1mo ago | |
| MMLU random topology | Inspector | ASR (k=1)15.5 | 16 | 15d ago | |
| URL-based PI (200-sample dataset) | ASR33.5 | 12 | 3mo ago | ||
| Spam Email | Separator Injection | ASR (None Defense)0.3 | 10 | 3mo ago | |
| Negative Review | Separator Injection | ASR (None Defense)0 | 10 | 3mo ago | |
| Toxic Comment | Topic Attack | ASR (None)100 | 10 | 3mo ago | |
| Prompt Overflow K=4 | DeBERTa Prompt v2 | Bypass Rate0 | 9 | 9d ago | |
| OpenClaw (140 adversarial instances) | ClawKeeper | Defense Success Rate90 | 7 | 2mo ago | |
| GCG Clean | CAHL | ASR37.02 | 4 | 3mo ago | |
| Representative guardrail dataset | ChainPoll | F1 Score97 | 3 | 3mo ago |