Vision language models show safety risks when reading scene states
Reading Scene States: Uncovering Safety Failures in Vision-Language Models
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Vision-language models can sometimes provide harmful advice by interpreting pictures in unsafe ways, even when not asked to. The authors found a new kind of vulnerability where the model’s understanding of a scene’s state leads to unsafe responses. They created a large test set to measure how often these models make mistakes and developed an attack method that tricks models into giving harmful answers. Their tests showed many models are vulnerable, and current safety filters are not enough to prevent these issues.
What this means in practice
- •For ai safety teams: Detect and mitigate safety risks in vision-language models by testing their scene-state interpretation abilities using SceneSafetyBench and S3TA methods.
- •For online content moderators: Improve moderation tools by understanding limitations of textual and visual filters against scene-state driven harmful responses in vision-language models.
Authors
Tongxi Wu, Yuyang Cheng, Jiacheng Xian, Menglin Yang, Kai Chen, Jian Zhang, Yang Gao
Abstract
Vision-language models (VLMs) can provide harmful assistance by interpreting visual events, even without an explicit request for the underlying harmful objective. We identify this \textbf{scene-state safety vulnerability} and introduce \textbf{SceneSafetyBench}, comprising 375 image--question pairs spanning Initial, Medium, and Ultimate scene states and five visual presentations. Across eight VLMs, scene-state inputs elicit harmful assistance with attack success rates (ASRs) of 45.6--72.3\%. Comparative analysis across three VLMs shows that scene-state inputs most closely resemble QR-related images in their layer-wise safety-related signal profiles, while exhibiting lower average late-layer readouts and distinct signal dynamics during generation. To turn this vulnerability into a systematic attack, we propose \textbf{S3TA}, which translates harmful objectives into scene states, queries image--question pairs independently, and assembles the responses. On seven MM-SafetyBench categories across eight open-weight and commercial VLMs, S3TA achieves a model-average ASR of \textbf{71.2\%, exceeding the strongest baseline by 6.1 percentage points}. Text-side filters and visual moderation provide limited protection. These findings highlight the need to safeguard the assistance models infer through scene-grounded reasoning, beyond screening input content.