GPU VulnDB

Database/AI/ML frameworks & serving

vLLM: clients can override the video frame cap via media_io_kwargs and exhaust API frontend memory

CVSS 8.7CVE-2026-116179AI/ML frameworks & servingcurated

Impact

The per-request media_io_kwargs field lets a caller set num_frames, overriding the server's frame-count cap. With num_frames=-1 and a large inline video, VideoMediaIO.load_base64 decodes every frame into memory and the API frontend dies. That takes out the process handling all requests for the replica, not just the offending one, so co-tenants on a shared inference endpoint lose service and the GPU workers behind it sit idle until the frontend is back. Distinct from the Nemotron-VL image-bomb issue (CVE-2026-116177): different code path, different affected range, different fix.

Who can reach it

Any API client able to POST a multimodal request to the vLLM frontend; the advisory describes no authentication requirement.

What to do

Upgrade vLLM to 0.31.0 or later and restart the API frontend and workers. As an interim mitigation, strip or reject media_io_kwargs at the gateway in front of vLLM and cap request body size.

References

Related entries

All AI/ML frameworks & serving entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.