mlsanityData to 7 Oct 2026

MLLM jailbreaks and adversarial robustness

850 papers9 this weekHigh confidence in this grouping: the papers form one clear topicHigh confidence in this grouping: the papers form one clear topic · About

Keywords: adversarial, safety, attacks, attack, multimodal

Key facts. MLLM jailbreaks and adversarial robustness: 850 papers on arXiv since 2022 (data through 7 Oct 2026). 166 of its 850 mapped papers (20%) cite “FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts” (2023, arXiv 2311.05608). 132 of its 850 mapped papers (16%) cite “Visual Adversarial Examples Jailbreak Aligned Large Language Models” (2023, arXiv 2306.13213). 27 new papers 17 Sep–7 Oct.

What’s new

17 Sep – 7 Oct 2026

27 new papers, 17 Sep – 7 Oct; 9 this week.

  1. This week · 9
  2. Detecting Adversarial Images through Response Profiles of Vision-Language Models
  3. Open-MMUnlearning: Unifying Methods and Evaluation for MLLM Unlearning also fits: MLLM training and visual instruction tuning
  4. Purifying Backdoored Large Vision-Language Models by Removing Hijacked Directions
  5. Understanding and Mitigating Token-Pruning-Induced Vulnerabilities in VLMs
  6. It Is Not Seeing the Hazard: A Frozen Vision-Language Safety Score Measures Its Caption Bank also fits: Driving VLMs and open-vocabulary detection
  7. Transferable Spatial Temporal Coherence Adversarial Attack on Black-Box Vision Language Models for Autonomous Driving also fits: Driving VLMs and open-vocabulary detection
  8. Reading, Not Manipulating: Leveraging Router Logits for Multimodal Safety in MoE Vision-Language Models
  9. From Local Evidence to Safety Verdicts: Causal Tracing in Vision-Language Models
  10. Visual Grounding Safety in Vision-Language Models
  11. Earlier · 18
  12. Visual-Invariance-Augmented Feature Optimal Alignment for Transferable Adversarial Attacks against Closed-Source MLLMs
All 27
  1. MLLMs Fail to Refuse when Using Tools Agentically
  2. Corrupted but Correct: Why Vision-Language Models Lie to Themselves Internally
  3. Typographic Attack Against VLM-based AI-generated Image Detection
  4. SceneJail: Exploiting Video Scenario Context to Jailbreak Multimodal LLMs
  5. UnifiedAttack: Evaluating the Safety of Large Multimodal Models in Synergistic Harmful Image-Text Generation
  6. It Takes Little to Rewrite Perception: Targeted Semantic Substitution in Vision-Language Models at ε≤ 4/255
  7. Can Vision-Language Models Stay Helpful When Facing Implicit Risks? Intent-Privilege OPSD for Efficient Safety-Helpfulness Alignment
  8. Selective Channel Restoration for Backdoored Vision-Language Models
  9. ThinkingGuard: Decoding Implicit Hazards via Step-by-Step Risk Attribution in Multimodal Large Language Models
  10. Look Before You Judge: Training-Free Region Mining for Grounded and Explainable Deepfake Detection also fits: VLM benchmarks and visual reasoning
  11. Still There, No Longer Seen: Exposing Compression-Induced Risk in Large Vision-Language Models also fits: Visual token pruning and efficiency
  12. VisionPsy-Nano: Improving Accuracy, Efficiency, and Reliability in On-Device Vision-Language Models also fits: VLM benchmarks and visual reasoning
  13. Evaluating the Semantic-to-Geometric Gap in Adversarial Defenses Against Vision-Language Model-Based Plagiarism
  14. A2Safe: Counterfactual Evidence-Aligned Adaptive Agent Collaboration for Safe and Effective Visual Question Answering also fits: VLM benchmarks and visual reasoning
  15. Hiding in Plain Sight: A Diffusion-based Mitigation of Geolocation Privacy Leakage in Vision-Language Models
  16. Fingerprinting Multimodal Large Language Models also fits: MLLM training and visual instruction tuning
  17. IMFD: End-to-end Multi-Face Forgery Detection through Instruction-based Large Vision-Language Models

Foundations

Most cited by papers here

  1. FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts2023cited by 166 here
  2. Visual Adversarial Examples Jailbreak Aligned Large Language Models2023cited by 132 here
  3. On Evaluating Adversarial Robustness of Large Vision-Language Models2023cited by 128 here
  4. MM-SafetyBench: A Benchmark for Safety Evaluation of Multimodal Large Language Models2023cited by 120 here
  5. Jailbreak in pieces: Compositional Adversarial Attacks on Multi-Modal Language Models2023cited by 105 here

Lineage

The papers this direction builds on, oldest first.

Each paper on the line cites the one before.

  1. 2022BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generationcited by 90 here
  2. 2022Flamingo: a Visual Language Model for Few-Shot Learningcited by 73 here
  3. 2023InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuningcited by 142 here
  4. 2023MME: A Comprehensive Evaluation Benchmark for Multimodal Large Language Modelscited by 44 here
  5. 2023MM-Vet: Evaluating Large Multimodal Models for Integrated Capabilitiescited by 49 here
  6. 2023MM-SafetyBench: A Benchmark for Safety Evaluation of Multimodal Large Language Modelscited by 120 here
  7. 2024AdaShield: Safeguarding Multimodal Large Language Models from Structure-based Attack via Adaptive Shield Promptingcited by 64 here
  8. 2024JailBreakV: A Benchmark for Assessing the Robustness of MultiModal Large Language Models against Jailbreak Attackscited by 64 here
16 more it builds on, not in this chain
  1. 2022Towards Adversarial Attack on Vision-Language Pre-training Modelscited by 69 here
  2. 2022Understanding Zero-Shot Adversarial Robustness for Large-Scale Modelscited by 53 here
  3. 2023On Evaluating Adversarial Robustness of Large Vision-Language Modelscited by 128 here
  4. 2023Visual Adversarial Examples Jailbreak Aligned Large Language Modelscited by 132 here
  5. 2023Set-level Guidance Attack: Boosting Adversarial Transferability of Vision-Language Pre-training Modelscited by 52 here
  6. 2023Jailbreak in pieces: Compositional Adversarial Attacks on Multi-Modal Language Modelscited by 105 here
  7. 2023Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyondcited by 66 here
  8. 2023How Robust is Google's Bard to Adversarial Image Attacks?cited by 74 here
  9. 2023Improved Baselines with Visual Instruction Tuningcited by 152 here
  10. 2023FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Promptscited by 166 here
  11. 2023InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Taskscited by 59 here
  12. 2024MLLM-Protector: Ensuring MLLM's Safety without Hurting Performancecited by 53 here
  13. 2024Robust CLIP: Unsupervised Adversarial Fine-Tuning of Vision Embeddings for Robust Large Vision-Language Modelscited by 52 here
  14. 2024Eyes Closed, Safety On: Protecting Multimodal LLMs via Image-to-Text Transformationcited by 61 here
  15. 2024Images are Achilles' Heel of Alignment: Exploiting Visual Vulnerabilities for Jailbreaking Multimodal Large Language Modelscited by 77 here
  16. 2024Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolutioncited by 86 here

The 24 most cited of 1,168.

Browse

30 of 850, ranked 6–35

  1. 6Images are Achilles' Heel of Alignment: Exploiting Visual Vulnerabilities for Jailbreaking Multimodal Large Language Models202477
  2. 7How Robust is Google's Bard to Adversarial Image Attacks?202374
  3. 8Towards Adversarial Attack on Vision-Language Pre-training Models202269
  4. 9AdaShield: Safeguarding Multimodal Large Language Models from Structure-based Attack via Adaptive Shield Prompting202464
  5. 10JailBreakV: A Benchmark for Assessing the Robustness of MultiModal Large Language Models against Jailbreak Attacks202464
  6. 11Eyes Closed, Safety On: Protecting Multimodal LLMs via Image-to-Text Transformation202461
  7. 12MLLM-Protector: Ensuring MLLM's Safety without Hurting Performance202453
  8. 13Understanding Zero-Shot Adversarial Robustness for Large-Scale Models202253
  9. 14Robust CLIP: Unsupervised Adversarial Fine-Tuning of Vision Embeddings for Robust Large Vision-Language Models202452
  10. 15Set-level Guidance Attack: Boosting Adversarial Transferability of Vision-Language Pre-training Models202352
All 30
  1. 16On the Robustness of Large Multimodal Models Against Image Adversarial Attacks202337
  2. 17Safety of Multimodal Large Language Models on Images and Texts202436
  3. 18Jailbreak Vision Language Models via Bi-Modal Adversarial Prompt202432
  4. 19Pre-trained Model Guided Fine-Tuning for Zero-Shot Adversarial Robustness202431
  5. 20VL-Trojan: Multimodal Instruction Backdoor Attacks against Autoregressive Visual Language Models202431
  6. 21A Survey of Attacks on Large Vision-Language Models: Resources, Advances, and Future Trends202431
  7. 22Abusing Images and Sounds for Indirect Instruction Injection in Multi-Modal LLMs202330
  8. 23White-box Multimodal Jailbreaks Against Large Vision-Language Models202429
  9. 24Red Teaming Visual Language Models202428
  10. 25One Prompt Word is Enough to Boost Adversarial Robustness for Pre-trained Vision-Language Models202428
  11. 26SPA-VL: A Comprehensive Safety Preference Alignment Dataset for Vision Language Model202428
  12. 27Adversarial Prompt Tuning for Vision-Language Models202328
  13. 28Shadowcast: Stealthy Data Poisoning Attacks Against Vision-Language Models202427
  14. 29Protecting Privacy in Multimodal Large Language Models with MLLMU-Bench202427
  15. 30A Survey of Safety on Large Vision-Language Models: Attacks, Defenses and Evaluations202525
  16. 31ImgTrojan: Jailbreaking Vision-Language Models with ONE Image202425
  17. 32ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time202424
  18. 33R-TPT: Improving Adversarial Robustness of Vision-Language Models through Test-Time Prompt Tuning202521
  19. 34Test-Time Backdoor Attacks on Multimodal Large Language Models202421
  20. 35An Image Is Worth 1000 Lies: Adversarial Transferability across Prompts on Vision-Language Models202421