CVE-2026-114055

vLLM request-level processor settings can exhaust a multimodal server

vLLM · Multimodal request processing

First documented disclosure · Review checkpoint

Summary

One inference request can override sizing parameters and cause excessive allocations on a shared service.

Severity and exploitation evidence

High · 8.7/10 · CVSS 4.0. Severity source.

Exploitation not established: The report documents a controllable allocation path. Malicious exploitation is not established.

Exploitation evidence source

Technical severity is not a measure of observed attacks or of risk to your deployment. Attempts, demonstrations, and successful compromise retain the qualifications in the source.

How the weakness works

The API forwards arbitrary processor parameters into model-specific preprocessing. Resizing images or increasing frame, patch or crop counts can amplify memory use.

Conditions required

An inference client can reach a model whose processor honors these overrides. Unprotected endpoints allow unauthenticated access.

Potential impact

CPU or GPU memory exhaustion can interrupt service for other clients. The effect depends on the processor and the resources assigned to the worker.

Affected versions and configurations

0.7.3 through versions before 0.31.0 per VulnCheck.

Fix and mitigation status

0.31.0 per vLLM. A trusted-client opt-in can reopen request-level overrides.

Practical next steps

Upgrade and keep client-selected multimodal processor options disabled on shared inference endpoints. Apply settings controlled by the operator on the server.

Operating lesson

Make expensive model-construction controls an operator decision instead of accepting arbitrary values from clients.

Evidence and limits

The September 26 public fix predates the October 6 advisory and October 21 CNA. This concerns resource use at the request boundary, distinct from remote code execution through code_revision.

Sources