NVIDIA CUDA Toolkit, nvJPEG library: Out-of-bounds read/write in nvJPEG while decoding an image, rated code-execution

Impact

Out-of-bounds read/write in nvJPEG while decoding an image, rated code-execution capable. nvJPEG sits in exactly the place you feed untrusted data: GPU-accelerated image decode in training and inference pipelines, DALI preprocessing, and media services. If your ingest path decodes user-uploaded JPEGs on the GPU, this is remote code execution reachable by whoever can upload an image.

Who can reach it

Anyone who can get a malformed JPEG into a pipeline that decodes with nvJPEG - typically an unauthenticated uploader, several hops upstream of the GPU.

What to do

Upgrade the CUDA Toolkit to 11.1.1 or later. The real cost is not the host install: every container image, wheel, conda package and vendored build that statically carries the affected library has to be rebuilt and re-pushed, and running jobs restarted onto the new image. No driver reload, no node drain, no firmware flash - but a full image-fleet rebuild.

References