CVE-2026-12491
Received Received - Intake
Improper Image Metadata Handling in vLLM

Publication date: 2026-06-17

Last updated on: 2026-06-17

Assigner: Red Hat, Inc.

Description
A flaw was found in vLLM, an open-source library for large language model inference. This vulnerability arises from improper handling of image metadata, specifically EXIF orientation and PNG transparency (tRNS) data, during image processing. When images are converted to RGB, transparency information may be implicitly discarded or remapped, leading to unexpected rendering of transparent pixels and distortion of input content. This can result in the model misinterpreting image content, potentially affecting the integrity of processed data.
CVSS Scores
EPSS Scores
Probability:
Percentile:
Meta Information
Published
2026-06-17
Last Modified
2026-06-17
Generated
2026-06-17
AI Q&A
2026-06-17
EPSS Evaluated
N/A
NVD
EUVD
Affected Vendors & Products
Currently, no data is known.
Helpful Resources
Exploitability
CWE
CWE Icon
KEV
KEV Icon
CWE ID Description
CWE-115 The product misinterprets an input, whether from an attacker or another product, in a security-relevant fashion.
Attack-Flow Graph
AI Quick Actions
Instant insights powered by AI
Executive Summary

CVE-2026-12491 is a vulnerability in the vLLM library, an open-source tool for large language model inference, caused by improper handling of image metadata during processing.

Specifically, the vulnerability arises because the EXIF orientation data in images is not normalized after loading, and PNG images with tRNS transparency are not properly flattened before conversion to RGB.

As a result, transparent or semi-transparent pixels may be rendered unexpectedly, causing the model to misinterpret the visual content of images, which can distort the input data.

Impact Analysis

This vulnerability can impact you by causing the vLLM model to misinterpret image content due to unexpected rendering of transparent pixels and distorted input data.

Such misinterpretation can affect the integrity of the processed data, potentially leading to incorrect outputs or decisions based on flawed image inputs.

The issue affects all Linux environments using the vLLM library and has a medium severity rating.

Detection Guidance

This vulnerability arises from improper handling of image EXIF orientation and PNG tRNS transparency in the vLLM library during image processing.

To detect if your system is affected, you should check if the vLLM library is installed and used in your environment, especially on Linux systems.

Since the issue relates to image processing, you can test by processing images with EXIF rotation or PNG transparency through vLLM and observe if the output images or model interpretations are distorted or incorrect.

There are no specific commands provided in the resources to detect this vulnerability directly on your network or system.

Mitigation Strategies

Immediate mitigation involves ensuring that the vLLM library properly normalizes image EXIF orientation and explicitly flattens PNG transparency before conversion to RGB.

Specifically, the vulnerability is caused by the absence of a call to ImageOps.exif_transpose to normalize EXIF orientation and improper handling of PNG tRNS transparency in non-RGBA modes.

Until a patch or update is applied, avoid processing images with EXIF rotation or PNG transparency through vLLM or preprocess images to normalize EXIF orientation and flatten transparency manually.

Monitor for updates or patches from the vLLM maintainers or your Linux distribution to apply fixes addressing this issue.

Chat Assistant
Ask questions about this CVE
Hi! I’m here to help you understand CVE-2026-12491. Ask me anything about the vulnerability, its impact, or mitigation strategies.
0/70
EPSS Chart