Research map / Vision-language and multimodal / Multimodal LLMs
Images are Achilles' Heel of Alignment: Exploiting Visual Vulnerabilities for Jailbreaking Multimodal Large Language Models
Yifan Li, hangyu guo, Kun Zhou, Wayne Xin Zhao, Ji-Rong Wen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add a blank image to a harmful text instruction | LLaVA-1.5 · Harmful-instruction dataset | average ASR (%): +25.20 | Hurt |
| replace blank image by a harmful (toxic) image | LLaVA-1.5 · Harmful-instruction dataset | average ASR (%): 54.13 -> 68.13 | Hurt |
| fine-tune more backbone parameters (full vs LoRA) | LLaVA-1.5 · Toxic setting | average ASR (%): 62.80 -> 68.13 | Hurt |
| add text-to-image pointer to typography image (ablation) | HADES · HADES dataset | average ASR (%): LLaVA-1.5 -7.47, LLaVA-1.5 LoRA -14.13, LLaVA +6.53, Gemini +16.93, GPT-4V +1.47 | Mixed |
| add optimized harmful image generated by diffusion with LLM prompt refinement (ablation) | HADES · HADES dataset | average ASR (%): LLaVA +33.86, Gemini Pro Vision +37.20, GPT-4V +13.34 | Improved |
| add gradient-optimized adversarial image on top (ablation) | HADES on LLaVA-1.5 · HADES dataset | average ASR (%): 78.53 -> 90.26 | Improved |
| more LLM-driven prompt optimization steps for harmful image generation (ablation) | HADES · HADES dataset | ASR | Improved |
| HADES instead of adversarial-image or compositional jailbreak baselines | LLaVA-1.5 · Violence/Animal/Self-Harm categories | ASR (%): Animal 74.67 -> 83.33; Self-Harm 80.67 -> 89.33 | Improved |
| fine-tune with contrastive harmlessness LoRA on OCR and captioning refusals | LLaVA-1.5 · HADES test split | average ASR (%): 89.53 -> 5.07 | Improved |
| add contrastive harmlessness LoRA | LLaVA-1.5 · LLaVA-Bench | score: 63.40 -> 57.90 | Hurt |
Closest papers by meaning
- MIDAS: Multi-Image Dispersion and Semantic Reconstruction for Jailbreaking MLLMs (ICLR 2026 · similarity 0.91)
- Distraction is All You Need for Multimodal Large Language Model Jailbreaking (CVPR 2025 · similarity 0.91)
- FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts (AAAI 2025 · similarity 0.91)
- Playing the Fool: Jailbreaking LLMs and Multimodal LLMs with Out-of-Distribution Strategy (CVPR 2025 · similarity 0.91)
- Jailbreaking Vision-Language Models Through the Visual Modality (ICML 2026 · similarity 0.90)
- Jailbreak in pieces: Compositional Adversarial Attacks on Multi-Modal Language Models (ICLR 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
