CVE-2026-105758

vLLM: Qwen2-VL / Qwen3-VL video samplers bound on request-controlled max_frames, which the num_frames ceiling does not reach

Published Oct 5, 2026 CVSS 3.1 · 5.3 Medium EPSS < 1% KEV · Not listed
Description
vLLM is an inference and serving engine for large language models. From 0.24.0 until 0.30.0, 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 ceilings. An unauthenticated caller can submit these values to the /tokenize endpoint, causing the sampler to decode every frame selected from attacker-controlled video input, consume disproportionate frontend memory, and potentially terminate the API process before scheduling or admission control. The Rust frontend is not affected because it rejects the media_io_kwargs field. This issue is fixed in version 0.30.0.

Analysis

AI-generated

Technical impact

The vulnerability is that vLLM versions 0.24.0 through 0.30.0 allow an attacker to specify unbounded max_frames and fps values in media_io_kwargs for the Qwen2VLVideoBackend and Qwen3VLVideoBackend when calling the /tokenize endpoint. This lack of a server‑side ceiling lets an unauthenticated caller trigger the decoder to process every attacker‑controlled video frame, consuming excessive frontend memory and possibly causing the API process to terminate before it can complete scheduling or admission control. The result is a denial‑of‑service condition, which is consistent with CWE‑770 (Uncontrolled Resource Consumption).

Affected Systems

Affected vendors and products: vllm‑project’s vllm inference engine. The issue is present in releases 0.24.0, 0.24.1, …, up to but not including 0.30.0; it was fixed with the release of v0.30.0. The Rust‑based frontend is not vulnerable because it rejects media_io_kwargs. Users running the Python or standard vllm frontends and deploying any of the affected Qwen2VLVideoBackend or Qwen3VLVideoBackend backends are at risk unless they upgrade.

Risk and Exploitability

The CVSS v3.1 score is 5.3, indicating moderate severity. No EPSS data is published, so the probability of exploitation is uncertain, and the vulnerability is not listed in CISA’s KEV catalog. An attacker can exploit the flaw simply by issuing an unauthenticated request to /tokenize with tailored media_io_kwargs values, making the vulnerability highly likely to be used if the service is exposed. The lack of user authentication and the direct nature of the endpoint make the attack vector straightforward. Even without authentication, the endpoint exposes a remote execution path that can drain resources in a way that could lead to process termination and service disruption.

Remediation

No solution or workaround provided in the CVE record.

OpenCVE Recommended Actions

  • Update vLLM to version 0.30.0 or later.
  • Validate or limit media_io_kwargs.video.max_frames and media_io_kwargs.video.fps on the server side.
  • Restrict access to the /tokenize endpoint to trusted users or network segments until the fix is applied.

Generated by OpenCVE AI on October 6, 2026 at 00:27 UTC.

Tracking

Sign in to view the affected projects.

Advisories

Source ID Title
Github GHSA Github GHSA GHSA-x6mc-67gf-chw4 vLLM: Qwen2-VL / Qwen3-VL video samplers bound on request-controlled max_frames, which the num_frames ceiling does not reach
History
Type Values Removed Values Added
Metrics ssvc

{'options': {'Automatable': 'yes', 'Exploitation': 'none', 'Technical Impact': 'partial'}, 'version': '2.0.3'}

Type Values Removed Values Added
First Time appeared Vllm
Vllm vllm
CPEs cpe:2.3:a:vllm:vllm:*:*:*:*:*:*:*:*
Vendors & Products Vllm
Vllm vllm
Type Values Removed Values Added
First Time appeared Vllm-project
Vllm-project vllm
Vendors & Products Vllm-project
Vllm-project vllm
Type Values Removed Values Added
Description vLLM is an inference and serving engine for large language models. From 0.24.0 until 0.30.0, 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 ceilings. An unauthenticated caller can submit these values to the /tokenize endpoint, causing the sampler to decode every frame selected from attacker-controlled video input, consume disproportionate frontend memory, and potentially terminate the API process before scheduling or admission control. The Rust frontend is not affected because it rejects the media_io_kwargs field. This issue is fixed in version 0.30.0.
Title vLLM: Qwen2-VL / Qwen3-VL video samplers bound on request-controlled max_frames, which the num_frames ceiling does not reach
Weaknesses CWE-770
References
Metrics cvssV3_1

{'score': 5.3, 'vector': 'CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:L'}

Weaknesses
  • CWE-770

    Allocation of Resources Without Limits or Throttling

Sources

Status
PUBLISHED
Assigner
GitHub_M
Published
Updated
Oct 8, 2026, 2:32 AM
Reserved
Oct 5, 2026, 7:11 PM
Status
Analyzed
Published
Oct 5, 2026, 11:17 PM
Modified
Oct 8, 2026, 3:16 AM
Severity
Moderate
Public date
Oct 5, 2026, 10:54 PM