vLLM: Request-selected PyNvVideoCodec GPU decode bypasses static VRAM reservation
cve-program · CVE-2026-69147
PublishedvLLM is an inference and serving engine for large language models. Prior to 0.28.0, request bodies for Chat Completions and Responses can set media_io_kwargs.video.video_backend to pynvvideocodec, and MediaConnector.fetch_video forwards that choice to VideoMediaIO even when startup configuration selected a software decoder. The engine's _reserve_mm_ipc_gpu_memory logic budgets decoder memory only from static configuration, so the request-selected VIDEO_LOADER_REGISTRY backend can create a CUDA context, decoder surfaces, and decoded-frame allocations that were not removed from the engine's KV-cache budget. An attacker able to submit video requests to a video-capable GPU deployment with PyNvVideoCodec installed can exhaust shared GPU memory, causing request failures, worker crashes, or denial of service. The first release containing the fix is version 0.28.0.
Read the publisher's advisoryAffected versions, risk and source history
Exploitation and risk
Source statusCISA KEV lists vulnerabilities known to be exploited. FIRST EPSS estimates the likelihood of exploitation in the next 30 days. An unavailable score does not mean zero risk.
Which CVEs have scores here?
This view includes predictions for CVE identifiers explicitly listed in public NuGet advisories. Microsoft release pages have their own prediction coverage. Check source status for the scope and latest checks.
No exploitation or prediction data is available here for these CVEs. This does not mean they are safe or unlikely to be exploited.
cve-program · CVE-2026-69147
PublishedvLLM: Request-selected PyNvVideoCodec GPU decode bypasses static VRAM reservation
vLLM is an inference and serving engine for large language models. Prior to 0.28.0, request bodies for Chat Completions and Responses can set media_io_kwargs.video.video_backend to pynvvideocodec, and MediaConnector.fetch_video forwards that choice to VideoMediaIO even when startup configuration selected a software decoder. The engine's _reserve_mm_ipc_gpu_memory logic budgets decoder memory only from static configuration, so the request-selected VIDEO_LOADER_REGISTRY backend can create a CUDA context, decoder surfaces, and decoded-frame allocations that were not removed from the engine's KV-cache budget. An attacker able to submit video requests to a video-capable GPU deployment with PyNvVideoCodec installed can exhaust shared GPU memory, causing request failures, worker crashes, or denial of service. The first release containing the fix is version 0.28.0.
- Source published
- 16 Sep 2026 17:49:20 UTC
- Source modified
- 16 Sep 2026 18:37:14 UTC
- Saved here
- 23 Oct 2026 15:48:27 UTC
- Source terms
- CVE Program Terms of Use
Original source record · Cite this exact observation · Complete original publisher JSON
Source: cve-program · CVE Program Terms of Use. The publisher's wording is preserved. ChangeIntel groups the information and labels changes. The JSON record includes additional source fields.
CVSS 3.1 · 6.5 CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H