CVE-2026-34755 — vLLM Affected by Denial of Service via Unbounded Frame Count in video/jpeg Base64 Processing

Description (CNA)

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.

Affected (CNA)

Vendor / product	Versions
vllm-project / vllm	affected >= 0.7.0, < 0.19.0

CWE

CWE-770

References

https://github.com/vllm-project/vllm/security/advisories/GHSA-pq5c-rjhq-qp7p

ADP enrichment

CISA-ADP

No English description supplied.

No references supplied.

redhat-SADP

A flaw was found in vLLM, an inference and serving engine for large language models. A remote attacker can exploit a vulnerability in the VideoMediaIO.load_base64() method by sending a single API request containing a large number of comma-separated base64-encoded JPEG frames. This bypasses the intended frame count limit, causing the server to decode all frames into memory. This can lead to an Out-of-Memory (OOM) crash, resulting in a Denial of Service (DoS) for the affected system.