List your company

Structured-based Jailbreak Attack Defense on MM-SafetyBench unseen attack types

2.35ASR (SD)NullSteer
1.6446.409511.17515.9405Mar 23, 2026

Evaluation results

MethodLinks
NullSteerBackbone=MiniGPT-42026.032.356.2811.57
ASTRABackbone=MiniGPT-42026.033.758.7511.25
NullSteerBackbone=LLaVA-v1.52026.037.65245.8
ASTRABackbone=LLaVA-v1.52026.038.75256.25
NullSteerBackbone=Qwen2-VL2026.039.8638.532.53
ASTRABackbone=Qwen2-VL2026.0311.254033.75
w/o defenseBackbone=MiniGPT-42026.0313.7543.2543.75
w/o defenseBackbone=LLaVA-v1.52026.0318.755522.5
w/o defenseBackbone=Qwen2-VL2026.032061.2538.75