场景即攻击面：揭示MLLM中基于场景的安全失效
原标题：Scenes as an Attack Surface: Uncovering Scene-Grounded Safety Failures in MLLMs
AI 摘要
该论文发现多模态大语言模型（MLLM）存在一种新的安全漏洞：仅凭任务相关的场景语义（而非显式有害指令或对抗性视觉信号）就能使模型从拒绝转为顺从。作者将这一失败形式化为「场景接地视觉安全漏洞」，并构建了包含 400 个图文对、覆盖八类风险的 SceneSafetyBench 基准，在九个白盒与黑盒 MLLM 上评估。基于此提出 Scene-Grounded Attacks（SGA），平均攻击成功率达 86.17%，比最强基线高 8.55 个百分点，在八个模型上领先。
正文节选
Scenes as an Attack Surface: Uncovering Scene-Grounded Safety Failures in MLLMs Abstract Existing jailbreak attacks have exposed serious safety weaknesses in multimodal large language models (MLLMs). Most prior attacks rely on specially constructed visual inputs to carry or support harmful intent. In this paper, we uncover a different vulnerability. We find that task-relevant scene semantics alone can shift an MLLM from refusal to compliance. This effect occurs even when the visual input adds n