CVE-2026-34755

vLLM is an inference and serving engine for large language models (LLMs). From 0.7.0 to before 0.19.0, the VideoMediaIO.load_base64() method at vllm/multimodal/media/video.py splits video/jpeg data URLs by comma to extract individual JPEG frames, but does not enforce a frame count limit. The num_frames parameter (default: 32), which is enforced by the load_bytes() code path, is completely bypassed in the video/jpeg base64 path. An attacker can send a single API request containing thousands of comma-separated base64-encoded JPEG frames, causing the server to decode all frames into memory and crash with OOM. This vulnerability is fixed in 0.19.0.

What this means for your business

It affects Vllm. It matters if your company, or a supplier that handles your data, runs it.
An attacker can use it remotely, over a network, with an ordinary user login, and without anyone at your company clicking anything.
FIRST's prediction model gives it a 0.5% chance of attack attempts being seen in the next 30 days, ranking above 39% of all known flaws.

What to do

1Check whether your company or your suppliers run Vllm, and which version. The affected versions are listed further down this page.
2If you do, apply the vendor's fix. A patch or vendor advisory has been published.

Not sure if your company is exposed?

Fastnexa’s certified penetration testers can check whether attackers could use this flaw, or others like it, against your websites, apps and network. The full test is free for our first 10 founding clients until 31 December 2026.

Check if you’re exposed Book a call

Scoring

CVSS: 6.5 (v3.1)
Vector: CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H
Weakness: CWE-770
Assigned by: security-advisories@github.com

Dates

Published: 2026-04-06
Last modified: 2026-08-25
Sources: NVD

Affected products

Vllm Vllm0.7.0 - 0.19.0

As listed in the NVD configuration data. Not a statement about your estate.