Executive brief
vLLM is a high-performance engine used to serve Large Language Models (LLMs) to users and applications. A vulnerability in how it processes video data allows an attacker to send a specially crafted request that consumes all available server memory. This can lead to a complete service crash, resulting in a denial of service for all users of the AI platform.
Technical details
The vulnerability exists in the VideoMediaIO.load_base64() method within vllm/multimodal/media/video.py. While the load_bytes() path correctly enforces a 'num_frames' limit (default 32), the load_base64() path for 'video/jpeg' MIME types uses an unbounded split operation on comma-separated data URLs. An attacker with network access and low privileges can submit a single API request containing thousands of base64-encoded JPEG frames. Each frame is decoded into a full numpy array in memory, leading to rapid memory exhaustion (OOM) and a process crash. This issue is tracked as CWE-770 and is resolved in version 0.19.0.
Affected products
- vllm-project vLLM >= 0.7.0, < 0.19.0
- Red Hat Red Hat AI Inference Server 3
- Red Hat Red Hat Enterprise Linux AI (RHEL AI) 3 3
- Red Hat Red Hat OpenShift AI (RHOAI)
Timeline
- 2026-04-03: advisory: GitHub Security Advisory published
- 2026-04-06: disclosed: CVE-2026-34755 published to NVD
- 2026-04-06: patched: Fix released in version 0.19.0