Executive brief
vLLM is an open-source library used for running large language models, including those that process images. A flaw in how it handles image metadata means that certain images (like those with specific rotation settings or transparency) may be misinterpreted by the AI. This could lead to the model making incorrect decisions or misidentifying content, potentially compromising the reliability of automated systems using this library.
Technical details
A vulnerability exists in vLLM's image processing pipeline due to the lack of normalization for EXIF orientation and improper handling of PNG tRNS transparency data. Specifically, the library fails to call ImageOps.exif_transpose after opening an image, causing a mismatch between human-perceived orientation and model input. Additionally, while RGBA-to-RGB conversion is handled, PNGs using tRNS in other modes (P/L/RGB) are converted to RGB via a path that discards or remaps transparency semantics. An attacker could provide specially crafted images to cause the model to misinterpret input content. This issue is similar to the 'AlphaDog' class of attacks where subtle overlay elements become visible or distort the input after incorrect flattening.
Affected products
- vLLM Project vLLM
- Red Hat Red Hat AI Inference Server 3
Timeline
- 2026-06-17: disclosed: Initial report via Red Hat and NVD