CVE-2026-105758

A MODERATE severity flaw was found in vLLM, as shipped in Red Hat AI Inference Server 3 and Red Hat OpenShift AI (RHOAI). The Qwen2VLVideoBackend and Qwen3VLVideoBackend classes accept request-level values for the media_io_kwargs.video.max_frames and media_io_kwargs.video.fps fields without enforcing server-side upper bounds. An unauthenticated remote attacker can submit requests with excessively high video frame counts or sampling rates to API endpoints such as /tokenize, causing the service to decode every frame selected from attacker-controlled video input. This consumes disproportionate frontend memory and can terminate the API process before scheduling or admission control, resulting in a denial of service with no impact on confidentiality or integrity (CVSS 5.3, AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:L, matching the upstream rating). The Rust frontend is not affected because it rejects the media_io_kwargs field. For the affected Red Hat AI Inference Server and RHOAI products the fix is deferred because the impact is MODERATE with a CVSS score below 7.0; the residual risk is limited to availability, where an unauthenticated remote attacker could crash the vLLM API process and make the inference service unavailable until it is restarted.