CVE-2026-115914
Awaiting Analysis Awaiting Analysis - Queue

Remote Code Execution in vLLM via Zero-Sum Image Path

Publication date: 2026-10-26

Last updated on: 2026-10-26

Assigner: GitHub, Inc.

Description
vLLM is an inference and serving engine for large language models. From 0.11.1 until 0.32.0, the DeepSeek-OCR, DeepSeek-OCR-2, and Unlimited-OCR image path allows a remote client that can reach the OpenAI-compatible API to submit a legal image whose normalized pixels sum to zero when that image is alone in its per-step encoder batch. _parse_and_validate_image_input returns no embedding, embed_multimodal propagates that value, and sanity_check_mm_encoder_outputs triggers a fatal assertion that shuts down the engine for all users. This issue is fixed in version 0.32.0.

CVSS Scores

EPSS Scores

Probability:
Percentile:

Meta Information

Published
2026-10-26
Last Modified
2026-10-26
Generated
2026-10-26
AI Q&A
2026-10-26
EPSS Evaluated
N/A
NVD
CVE-2026-115914
EUVD
EUVD-2026-110032

Affected Vendors & Products

Showing 1 associated CPE
Vendor Product Version / Range
vllm-project vllm >= 0.11.1, < 0.32.0

Helpful Resources

Exploitability

CWE
CWE Icon
KEV
KEV Icon
CWE ID Description
CWE-617 The product contains an assert() or similar statement that can be triggered by an attacker, which leads to an application exit or other behavior that is more severe than necessary.

Attack-Flow Graph

AI Quick Actions

Instant insights powered by AI
Executive Summary

CVE-2026-115914 is a vulnerability in vLLM affecting deployments using DeepSeek-OCR models. It allows a remote attacker to crash the entire vLLM engine by sending a specific PNG image. The image causes the engine's encoder to return no embedding, triggering a fatal assertion that shuts down the engine for all users. The issue occurs when the sum of normalized pixel values in an image is exactly zero, which can happen with a legitimate half-black, half-white 1024x1024 PNG image.

The vulnerability stems from a flawed input validation check in the DeepSeek-OCR model that returns None for zero-sum pixel tensors. This violates the expected return type and causes the engine to fail. The issue was introduced in 2025 and affects versions 0.11.1 through 0.32.0.

Detection Guidance

To detect this vulnerability, monitor vLLM engine logs for fatal assertion errors or crashes triggered by image processing. Check for repeated API server unresponsiveness after processing PNG images. Use network traffic analysis to identify requests containing 1024x1024 PNG files with half black and half white pixels.

Impact Analysis

This vulnerability allows an unauthenticated attacker to crash the vLLM engine with a single request, making the API server unresponsive until manually restarted. The attack requires no special privileges or user interaction and can be repeated after each restart. It takes approximately 0.1 seconds per request and works even on idle instances or with retries.

Mitigation Strategies

Upgrade vLLM to version 0.32.0 or later. If upgrading is not immediately possible, disable DeepSeek-OCR, DeepSeek-OCR-2, and Unlimited-OCR model support in your vLLM deployment. Implement rate limiting on API endpoints to reduce attack surface.

Chat Assistant

Ask questions about this CVE
Hi! I’m here to help you understand CVE-2026-115914. Ask me anything about the vulnerability, its impact, or mitigation strategies.
0/70

EPSS Chart