Hard to Read, Easy to Jailbreak: How Visual Degradation Bypasses MLLM Safety Alignment
Zhixue Song, Boyan Han, Yiwei Wang, Chi Zhang
48.00/100
🫥 Mediocre
Incremental, thin
Content 48.0 · Citation bonus +0.0 · no citation data
💡 This paper reveals that MLLMs' safety alignment breaks down when processing low-resolution visually compressed text, proposes a cognitive overload explanation and a decoupled pipeline to mitigate the
#多模态安全#越狱攻击#视觉压缩#认知过载#对齐失效#multimodal safety#jailbreak attack#visual compression#cognitive overload#alignment failure
Score breakdown
Novelty6.0 / 10
Rigor5.0 / 10
Significance8.0 / 10
Clarity7.0 / 10
Reproducibility4.0 / 10
This tone hasn't been generated yet — roast it again to create it.