List your company

Prompt-level trigger detection on HarmBench DirectRequest

57.5Trigger Coverage @8AERIC
0.622415.388730.15544.9213May 13, 2026

Evaluation results

MethodLinks
AERICTarget Model=Qwen3-8B, Safe-trigger budget=10%2026.0557.558.7561.2564.3832.5469
AERICTarget Model=Gemma-4-E4B-it, Safe-trigger budget=10%2026.0534.3835.9440.6246.5623.5281
WildGuard-7B PrefixTarget Model=Gemma-4-E4B-it, Protocol=Prefix2026.0522.1925.3128.7537.515.3313
WildGuard-7B PrefixTarget Model=Qwen3-8B, Protocol=Prefix2026.0513.1314.3814.3814.387.3125
Qwen3Guard-Stream-4BTarget Model=Qwen3-8B2026.054.695.318.1310.634.3094
Qwen3Guard-Stream-4BTarget Model=Gemma-4-E4B-it2026.054.387.515.3132.819.3375
ShieldGemma-9B PrefixTarget Model=Qwen3-8B, Protocol=Prefix2026.053.127.8116.5621.886.7625
ShieldGemma-9B PrefixTarget Model=Gemma-4-E4B-it, Protocol=Prefix2026.052.816.8814.6920.946.3937