Reproduce “Hard to Read, Easy to Jailbreak: How Visual Degradation Bypasses MLLM Safety Alignment” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhixue Song, Boyan Han, Yiwei Wang, Chi Zhang

Researchers at Westlake University identified a novel security flaw in Multimodal Large Language Models (MLLMs) where visually degrading input text (e.g., reducing image resolution) paradoxically increases the success rate of harmful prompts, coining this the "Attack Comfort Zone" (ACZ). This vulnerability, observed across 10 state-of-the-art MLLMs, arises from a "visual cognitive overload" that delays safety feature activation, prompting the development of a "Structured Cognitive Offloading" defense that effectively neutralizes the ACZ by enforcing a serialized transcription and safety audit pipeline.

Key insights

  • MLLMs exhibit an "Attack Comfort Zone" (ACZ), an inverted-U shaped vulnerability curve where moderate visual degradation (e.g., 45-150 DPI) significantly increases the Attack Success Rate (ASR) of harmful prompts, despite high text legibility.
  • This ACZ phenomenon is caused by a "safety-feature delay," where early MLLM layers, overloaded by visual processing, defer the identification of harmful content to deeper layers, allowing malicious instructions to bypass initial safety guardrails.
  • The proposed "Structured Cognitive Offloading" defense, by explicitly decoupling visual transcription from safety assessment, effectively restores safety alignment for degraded inputs without compromising the MLLM's benign utility.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.