CVE-2026-105760: vLLM: GLMGA video sampling permits request-driven CPU and memory exhaustion

ChatCompletionRequest exposes media_io_kwargs as request-controlled values, and the connector merges them over configured defaults, including video_backend. A caller can therefore select the glmga backend even when the served model does not use it. GLMGA computes extract_t = min(int(duration * fps), max_frames), builds a Python list with that many entries and deduplicates it before the decoder reads anything.

The proof of concept sends an inline two-frame video with "video_backend": "glmga", "backend": "opencv", "fps": 500000 and "max_frames": 500000. The decoder ends up with two frames, but the intermediate list is huge. That work runs in the shared media-loading executor, so it delays unrelated media requests.

The advisory claims partial denial of service only, with no memory corruption, disclosure or code execution. It is reachable without credentials when no API key is set, and by any key holder otherwise. Existing byte, pixel or decoded-frame limits do not bound this pre-decode allocation.

vllm >= 0.23.0rc2, < 0.30.0, on servers that expose chat completions for a video-capable model and accept request-level media_io_kwargs. The model does not need to use GLMGA by default. No GPU decoder or external media host is needed to trigger it.

pip show vllm

Upgrade to vLLM 0.30.0:

pip install -U vllm==0.30.0

The fix (PR #54935) caps GLMGA sampling work. The advisory also lists workarounds:

• Filter request-level video_backend, fps and max_frames at the gateway.
• Do not let untrusted callers select GLMGA.
• Add authentication, rate limits, concurrency limits and memory isolation.
• Use a separate, constrained media-loading worker pool where possible.

Search request logs for media_io_kwargs bodies with "video_backend": "glmga" or very large fps and max_frames values such as 500000. On the host, look for the API-server process spiking in CPU and memory, with media requests stalling while small videos are submitted. The sources give no specific log line.

Summary

The OpenAI-compatible chat endpoint accepts request-level video loader options through media_io_kwargs. A caller can select the GLMGA sampler and provide large fps and max_frames values. GLMGA first constructs an attacker-sized Python index list and then deduplicates it, even when the supplied video contains only a few frames. This work occurs in the shared media-loading executor before frame reads, allowing a tiny valid video and compact JSON options to consume disproportionate CPU time and memory and delay unrelated media requests.

The demonstrated impact is partial denial of service. No memory corruption, data disclosure, or code execution is claimed.

Affected configuration

The server must expose chat completions for a video-capable model and accept request-level media_io_kwargs. The model does not need to use GLMGA by default because the request value overrides the model's loader mapping. The proof of concept uses an inline two-frame video and explicitly selects the OpenCV decoder, so no external media host or GPU decoder is required.

The request is reachable without vLLM credentials when no API key is configured. With API-key authentication enabled, any caller holding a valid key can reach the same path.

Attack surface

A remote caller submits a valid chat-completion request with a small video and the following request-level options:

{
  "media_io_kwargs": {
    "video": {
      "video_backend": "glmga",
      "backend": "opencv",
      "fps": 500000,
      "max_frames": 500000
    }
  }
}

Increasing both numeric values increases temporary list construction and deduplication work even when the final decoded frame set remains unchanged.

Root cause

1. ChatCompletionRequest exposes media_io_kwargs as request-controlled nested values: `protocol.py#L365-L371`.
2. The request options are carried into chat parameters without a numeric work bound on GLMGA's fps or max_frames: `protocol.py#L571-L599`.
3. Media options are merged so request values override configured defaults, including video_backend: `connector.py#L577-L605`.
4. The connector selects a registered video loader and runs media loading through the shared executor: `video.py#L28-L66` and `connector.py#L44-L47`.
5. GLMGA calculates extract_t = min(int(duration * fps), max_frames), builds a list with that many entries, and deduplicates it before decoding frames: `video.py#L667-L740`.
6. The OpenCV decoder receives the already-deduplicated indices, so a two-frame file can trigger a large intermediate allocation while producing only two decoded frames: `opencv.py#L28-L50`.

The missing invariant is a strict upper bound on sampling work before the index list is constructed. Bounding decoded output after deduplication does not bound the vulnerable intermediate computation.

Suggested remediation

Validate request-level video sampling options before dispatching work to the media executor. Enforce conservative absolute limits for fps, max_frames, and especially the computed candidate count. Reject non-finite, negative, or otherwise invalid numeric values.

Avoid constructing O(extract_t) intermediate Python lists. Generate bounded unique frame indices directly from the source frame count and output-frame limit. Apply the limit before list construction and before shared executor submission. Add tests showing that extreme request values are rejected or consume constant memory for a fixed output-frame count.

Workarounds

• Remove or filter request-level video_backend, fps, and max_frames options at the gateway.
• Do not allow untrusted callers to select GLMGA.
• Apply authentication, rate limiting, request concurrency limits, and process memory isolation.
• Use a separate constrained media-loading worker pool where operationally feasible.
• Existing media byte, pixel, or decoded-frame limits do not necessarily bound this pre-decode candidate-index allocation.

Related advisory

GHSA-cqm8-jxg6-fqfq concerns partial denial of service in the DeepStream video backend through backend confusion and insufficient pixel guarding. This report concerns a different CPU-side GLMGA algorithm: request-controlled sampling values create a large redundant index list before the OpenCV decoder reads a fixed, tiny frame set. The vulnerable backend, resource sink, and required remediation differ.