Megadose The AI curve, compressed.

Reading Scene States: Uncovering Safety Failures in Vision-Language Models

· ArXiv · AI/CL/LG ·
The researchers’ S3TA attack achieved a 71.2% average success rate across eight vision-language models, beating the strongest baseline by 6.1 percentage points.

The paper finds that models can provide harmful assistance by interpreting visual scenes without an explicit request for the harmful objective. Its SceneSafetyBench contains 375 image–question pairs covering three scene stages and five visual presentations. S3TA turns harmful objectives into scene states, queries image–question pairs independently, and assembles the responses. The authors report that text-side filters and visual moderation offer limited protection.

ArXiv · AI/CL/LG's note

score 5

Categories: Research