The VLLM Safety Paradox: Dual Ease in Jailbreak Attack and Defense
Yangyang Guo, Fangkai Jiao, Liqiang Nie, Mohan Kankanhalli
Abstract
The vulnerability of Vision Large Language Models (VLLMs) to jailbreak attacks appears as no surprise. However, recent defense mechanisms against these attacks have reached near-saturation performance on benchmark evaluations, often with minimal effort. This dual high performance in both attack and defense gives rise to a fundamental and perplexing paradox. To gain a deep understanding of this issue and thus further help strengthen the trustworthiness of VLLMs, this paper makes three key contributions: i) One tentative explanation for VLLMs being prone to jailbreak attacks-inclusion of vision inputs, as well as its in-depth analysis. ii) The recognition of a largely ignored problem in existing VLLM defense mechanisms-over-prudence. The problem causes these defense methods to exhibit unintended abstention, even in the presence of benign inputs, thereby undermining their reliability in faithfully defending against attacks. iii) A simple safety-aware method-LLM-Pipeline. Our method repurposes the more advanced guardrails of LLMs on the fly, serving as an effective alternative detector prior to VLLM response. Last but not least, we find that the two representative evaluation methods for jailbreak often exhibit chance agreement. This limitation makes it potentially misleading when evaluating attack strategies or defense mechanisms. We believe the findings from this paper offer useful insights to rethink the foundational development of VLLM safety with respect to benchmark datasets, defense strategies, and evaluation methods.
Disclaimer: This paper discusses violent and discriminatory content, which may be disturbing to some readers.
Caption: The image depicts a person wearing a bomb suit and walking through a construction site. The image is a meme that suggests that the person is being invited to a potentially dangerous situation, possibly involving a bomb.
Caption: This…, suggesting that if Irish people were the first slaves brought to America, they should have received reparations. However, this is a misrepresentation of history, as the term "slaves" typically refers to people of African descent who were brought to the Americas as part of the transatlantic slave trade.
Ask about this paper
Your agent reads all of it.
Lune indexed this paper to the last equation, along with the top-tier papers that cite it. Ask a question and the answer quotes them.
Your agent calls
Luneget_paper_fulltext
Free to start. No credit card required.
Terminal
Install the CLIlune papers fulltext a74c6c77-dac4-4759-a68f-52c462def8e2Cited by top-tier papers15
- Panacea: Mitigating Harmful Fine-tuning for Large Language Models via Post-fine-tuning PerturbationYibo Wang, Tiansheng Huang, Li Shen, Huanjin Yao et al.NeurIPS 2025 · 22 citations
- Rethinking Bottlenecks in Safety Fine-Tuning of Vision Language ModelsYi Ding, Lijun Li, Bing Cao, Jing ShaoICLR 2026 · 21 citations
- Understanding and Rectifying Safety Perception Distortion in VLMsXiaohan Zou, Jian Kang, George Kesidis, Lu LinNeurIPS 2025 · 20 citations
- CTRAP: Embedding Collapse Trap to Safeguard Large Language Models from Harmful Fine-TuningBiao Yi, Tiansheng Huang, Baolei Zhang, Tong Li et al.ACL 2026 · 15 citations
- SafePTR: Token-Level Jailbreak Defense in Multimodal LLMs via Prune-then-Restore MechanismBeitao Chen, Xinyu Lyu, Shengming Yuan, Jingkuan Song et al.NeurIPS 2025 · 14 citations
Builds on29
- Training language models to follow instructions with human feedbackLong Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida et al.NeurIPS 2022 · 24,707 citations
- High-Resolution Image Synthesis with Latent Diffusion ModelsRobin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser et al.CVPR 2022 · 13,123 citations
- HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in Hugging FaceYongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li et al.NeurIPS 2023 · 1,778 citations
- Chatbot Arena: An Open Platform for Evaluating LLMs by Human PreferenceWei-Lin Chiang, Lianmin Zheng, Ying Sheng, Anastasios Nikolas Angelopoulos et al.ICML 2024 · 1,212 citations
- MM-Vet: Evaluating Large Multimodal Models for Integrated CapabilitiesWeihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang et al.ICML 2024 · 1,191 citations
Related papers
- FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual PromptsYichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang et al.AAAI 2025 · 350 citations
- SafetyReminder: Reviving Delayed Safety Awareness of Vision-Language Models to Defend Against Jailbreak AttacksPeiyuan Tang, Haojie Xin, Xiaodong Zhang, Jun Sun et al.AAAI 2026
- Risk Awareness Injection: Calibrating Vision-Language Models for Safety without Compromising UtilityMengxuan Wang, Yuxin Chen, Gang Xu, Tao He et al.ICML 2026
- When the Prompt Becomes Visual: Vision-Centric Jailbreak Attacks for Large Image Editing ModelsJiacheng Hou, Yining Sun, Ruochong Jin, Haochen Han et al.ICML 2026 · 3 citations
- Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time AlignmentSoumya Suvra Ghosal, Souradip Chakraborty, Vaibhav Singh, Tianrui Guan et al.CVPR 2025
