Reproduce “PolyJailbreak: Cross-Modal Jailbreaking Attacks on Black-Box Multimodal LLMs” 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

Xinkai Wang, Beibei Li, Zerui Shao, Ao Liu, Guangquan Xu, Shouling Ji

This research identifies "multimodal safety asymmetry" as a core vulnerability in Multimodal Large Language Models (MLLMs), where visual alignment compromises safety mechanisms. PolyJailbreak, a black-box framework, leverages this asymmetry through a reinforcement learning approach to achieve over 95% attack success rates on leading MLLMs like GPT-4o and Gemini, demonstrating robust and transferable jailbreak capabilities.

Key insights

  • Visual alignment processes in MLLMs, particularly trainable backbone alignments, introduce 'multimodal safety asymmetry' by interfering with text-based safety mechanisms and creating weaker visual safety constraints.
  • Visual inputs, even semantically blank or noisy, consistently reduce the Cluster Separation Ratio (CSR) between benign and malicious hidden state representations, acting as latent triggers and amplifiers of jailbreak vulnerabilities.
  • MLLM defenses are vulnerable to a wide and diverse spectrum of multimodal attack pathways, as successful jailbreaks frequently result from varied combinations of attack strategies, not just a few dominant ones.

Summary by alphaXiv.

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