Skip to content

Translation in progress Some entries still include text in the original language.

Oct 12 · Monday

All papers

Found 9 papers
Filter3

Attack type

Jailbreaks9

Attack technique

Gradient and representation optimization5Automated search22Multi-turn escalation4Encoding and obfuscation19Role-play and persuasion6Conversation structure manipulation3Reasoning-chain manipulation3Cross-modal payloads9Defense-targeted24Dormant triggers21

Impact

Harmful content9

System type

Models and APIs8

Attacker access

Black-box9

Attack surface / entry point

User input8Images, audio and video9

Layer

Prompt layer8

System component

Guardrails and judges9Reasoning chains3Fine-tuning and open weights0Vision47Audio7Image generation4Multimodal57
Select any value within a category; different categories must all match. Counts include the other filters you have selected.Another 265papers have no tags yet and are not included in filtered results.
Attack technique：Cross-modal payloadsSystem component：Guardrails and judgesSystem component：Fine-tuning and open weightsClear all

Another 265papers do not have category tags yet and are not included in filtered results.

  1. AttacksarXiv 2609.07216

    DIVA：利用视觉锚点对视频生成模型实施多模态越狱

    提出 DIVA，将有害意图拆成视觉锚点与运动文本以越狱视频生成模型

    arXiv
    2609.07216
    Published
    Layer
    Prompt layer
    Setting
    Models and APIs
    Attacker
    Black-box
    Evaluated models
    Wan, CogVideoX, LTX-Video and others9
    AbstractDIVA 用视觉锚点抑制语义漂移做视频越狱，并给出多条件安全基准 TI2VSafetyBench。

    DIVA 是针对带参考图的视频生成模型的免训练 black-box 越狱：用一致性机制本身堵住语义漂移这条安全逃逸路径。。

    Paper analysisRead analysis
  2. AttacksarXiv 2606.01837

    论文提出 DSR 跨模态越狱框架：用良性输入诱导 MLLM 有害输出

    提出 DSR 跨模态越狱，用无害图文诱导 MLLM 生成有害图像

    arXiv
    2606.01837
    Published
    Layer
    Prompt layer
    Setting
    Models and APIs
    Attacker
    Black-box
    Evaluated models
    gemini-2.5-flash-image, qwen-image-2.0, Tongyi Wanx 2.5
    AbstractDSR 用无害图文诱导商业 MLLM 输出有害图像，作者称组合能力会反过来扩大攻击面。

    DSR 是一种 black-box 跨模态越狱：输入侧不放显式有害内容，而让商业 MLLM 在推理中把无害图文重组成有害图像。

    Paper analysisRead analysis
  8. AttacksarXiv 2511.16110

    MFA 框架在 17 个视觉语言模型上取得 58.5% 攻击成功率

    提出 MFA 框架，协同突破视觉语言模型的对齐、提示与审核层

    arXiv
    2511.16110
    Published
    Layer
    Prompt layer
    Setting
    Models and APIs
    Attacker
    Black-box
    Evaluated models
    MiniGPT-4, LLaMA-4-Scout-Inst, LLaMA-3.2-11B-Vision-Instruct and others17

    Abstract作者称多层防御的 VLM 仍存在跨模型漏洞，MFA 整合三项技术实现了端到端越狱。

    Paper analysisRead analysis
  9. AttacksarXiv 2503.06223

    RedDiffuser：用强化扩散生成视觉输入审计 VLM 多模态安全失效

    提出 RedDiff，用强化扩散生成图像审计 VLM 情境安全失效

    arXiv
    2503.06223
    Published
    Layer
    Prompt layer
    Setting
    Models and APIs
    Attacker
    Black-box
    Evaluated models
    LLaVA-1.5-7B, Gemini-1.5-flash, LLaMA-3.2-11B-Vision-Instruct
    Abstract用强化扩散生成视觉上下文，审计有害文本情境下的 VLM 安全失效。

    RedDiffuser 是黑盒审计框架：有害内容不是显式攻击指令，而是固定的部分有毒文本，再配上扩散模型生成的图像，看 VLM 的开放续写会不会变得不安全。

    Paper analysisRead analysis
You have reached the end