CVE-2026-116179: vLLM: num_frames override enables memory-exhaustion DoS

vLLM versions 0.18.0 through 0.30.x let any API client override the video frame-count cap via media_io_kwargs, so a single request with num_frames set to -1 and a large inline video forces the server to decode every frame, exhaust memory and crash the API frontend. The attack needs no authentication, no user interaction and is low complexity (CVSS 7.5, availability impact only), so any internet-reachable multimodal vLLM endpoint is exposed. There is no EPSS score, no public exploit, no Nuclei template and no KEV listing at publication, so active exploitation is not currently evidenced, but the trivial request shape means weaponization would be fast. Upgrade to vLLM 0.31.0 or later, and until then block or strip media_io_kwargs at your gateway and cap request body sizes.

No confirmed active exploitation of CVE-2026-116179 has been reported, but organizations should still patch proactively.

1) Upgrade vLLM to 0.31.0 or later. 2) If patching is delayed, filter or reject the media_io_kwargs field (especially num_frames) at a reverse proxy or API gateway, or disable video input if not needed. 3) Enforce maximum request body size and per-client rate limits. 4) Require authentication (API key or mTLS) on vLLM endpoints and do not expose them directly to the internet. 5) Set container and process memory limits with automatic restart, and run multiple replicas behind a load balancer. 6) Detect: alert on OOM kills, frontend restarts, abnormally large request bodies, and requests containing num_frames with negative or very large values. Review the GitHub advisory GHSA-4m6g-h9mv-wqjc for details.

This vulnerability affects the following AI/ML architecture patterns: model serving, LLM inference APIs, multimodal/video inference pipelines, agent frameworks relying on hosted inference.

CVE-2026-116179 has a CVSS v3.1 base score of 7.5 (HIGH). The EPSS exploitation probability is 0.32%.