vLLM interpretation conflict in image EXIF and transparency handling

Executive brief

vLLM is a high-performance engine used for serving large language models, including those that process images. A vulnerability in how it handles image files means that the AI model may 'see' a different version of an image than what a human user intended, due to incorrect handling of image rotation and transparency. This could lead the AI to make incorrect decisions or reveal hidden information in images that were supposed to be transparent, potentially impacting the reliability and safety of AI-driven operations.

Technical details

vLLM contains an interpretation conflict vulnerability (CWE-436) in its multimodal image processing pipeline. Specifically, the library fails to call `ImageOps.exif_transpose` after opening images, causing the model to process images in their raw orientation rather than the human-intended EXIF-defined orientation. Additionally, while RGBA to RGB conversion is handled correctly, other modes such as P, L, or RGB with tRNS chunks are not explicitly flattened against a background; instead, they are converted directly to RGB, which can cause transparent or semi-transparent pixels to render as solid colors. This allows for 'AlphaDog' style attacks where hidden content becomes visible to the model, potentially altering downstream reasoning. A fix has been merged in vLLM pull request #44974.

Affected products

• vLLM Project vLLM >= 0.11.0, <= 0.23.0

Timeline

• 2026-06-11: disclosed: Initial disclosure to vLLM project
• 2026-06-17: advisory: GitHub Advisory published

References

• https://api.github.com/users/kexinoh
• https://github.com/kexinoh
• https://api.github.com/users/kexinoh/gists%7B/gist_id%7D
• https://api.github.com/users/kexinoh/repos
• https://avatars.githubusercontent.com/u/91727108?v=4
• https://api.github.com/users/kexinoh/events%7B/privacy%7D