Safety Classifier Evasion on SND (Secret-Note Diffusion) filter-evasion set
100∂ins Success RateLlamaGuard3 (F1)
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| LlamaGuard3 (F1)Developer=Meta, Training scope=13 general harm categories (MLCC taxonomy)2026.05 | 100 | 100 | 100 | 100 | 80 | 95.9 | |
| AprielGuard (F3)Developer=ServiceNow AI (SLAM Lab), Training scope=Memory poisoning + agentic exploits specifically; arXiv:2512.202932026.05 | 100 | 100 | 100 | 100 | 100 | 100 | |
| Granite Guardian (F4)Developer=IBM, Training scope=Harm + jailbreak detection (enterprise)2026.05 | 100 | 100 | 100 | 100 | 100 | 100 | |
| WildGuard (F2)Developer=AllenAI, Training scope=Adversarial prompts (WildGuard test set)2026.05 | 85.7 | 100 | 100 | 95.5 | 70 | 90 |