List your company
ResearchFormerly WizWand

Adversarial Detection on Direct Harmful Prompts

100DSRGradSafe
-4235077May 2, 2026
Updated 6mo ago

Evaluation results

MethodLinks
GradSafeTarget Model=Llama-3.1-8B-Instruct, Grad=true2026.05100
GradSafeTarget Model=Qwen2.5-7B-Instruct, Grad=true2026.0599.4
Linear ProbeTarget Model=Qwen2.5-7B-Instruct, Params=<1M, Grad=false2026.0599
SALOTarget Model=Qwen2.5-7B-Instruct, Params=<20M, Grad=false2026.0599
SALOTarget Model=Mistral-7B-Instruct-v0.3, Params=<20M, Grad=false2026.0599
Linear ProbeTarget Model=Mistral-7B-Instruct-v0.3, Params=<1M, Grad=false2026.0598.7
SALOTarget Model=Llama-3.1-8B-Instruct, Params=<20M, Grad=false2026.0598.3
No Defense (1-ASR)Target Model=Qwen2.5-7B-Instruct2026.0597.9
No Defense (1-ASR)Target Model=Llama-3.1-8B-Instruct2026.0596.9
Smooth LLMTarget Model=Qwen2.5-7B-Instruct, Grad=false2026.0594.4
Linear ProbeTarget Model=Llama-3.1-8B-Instruct, Params=<1M, Grad=false2026.0593.4
GradSafeTarget Model=Mistral-7B-Instruct-v0.3, Grad=true2026.0588.1
Smooth LLMTarget Model=Llama-3.1-8B-Instruct, Grad=false2026.0581.7
No Defense (1-ASR)Target Model=Mistral-7B-Instruct-v0.32026.0565
Smooth LLMTarget Model=Mistral-7B-Instruct-v0.3, Grad=false2026.0558.1
PPL FilterTarget Model=Mistral-7B-Instruct-v0.3, Grad=false2026.051.5
PPL FilterTarget Model=Qwen2.5-7B-Instruct, Grad=false2026.050
PPL FilterTarget Model=Llama-3.1-8B-Instruct, Grad=false2026.050