vLLM Denial of Service via Unbounded Video Sampler Frames

# vLLM Denial of Service via Unbounded Video Sampler Frames

https://gcdradar.vercel.app/incidents/vllm-denial-of-service-via-unbounded-video-sampler-frames · AIR-2026-0240 · closed 2026-10-05

- Lane: Patch now
- Severity: 2 (Disclosed vulnerability)
- Fix status: Fix status unknown
- First reported: 2026-10-05
- AI role: Attack on AI
- Products: vLLM
- CVE ids: CVE-2026-105758
- CVSS: 5.3

## Summary

An unauthenticated caller can submit malicious video frame parameters to the vLLM tokenizer endpoint, consuming excessive frontend memory and crashing the API process. The vulnerability is addressed in updated framework releases.

## What to do

No action recorded; read the sources below and the vendor's advisory.

## Latest updates

- 2026-10-05: CVE-2026-105758: vLLM: Qwen2-VL / Qwen3-VL video samplers bound on request-controlled max_frames, which the num_frames ceiling does not reach (https://github.com/advisories/GHSA-x6mc-67gf-chw4)
- 2026-10-05: CVE-2026-105758: vLLM is an inference and serving engine for large language models. (https://nvd.nist.gov/vuln/detail/CVE-2026-105758)

## Sources

- 2026-10-05 NVD: CVE-2026-105758: vLLM is an inference and serving engine for large language models. https://nvd.nist.gov/vuln/detail/CVE-2026-105758
- 2026-10-05 GitHub Advisory Database: CVE-2026-105758: vLLM: Qwen2-VL / Qwen3-VL video samplers bound on request-controlled max_frames, which the num_frames ceiling does not reach https://github.com/advisories/GHSA-x6mc-67gf-chw4

Verify against the vendor's advisory before changing production. Data CC BY-SA 4.0, https://gcdradar.vercel.app.