CVE-2026-126398
Awaiting Analysis Awaiting Analysis - Queue

vLLM GPUModelRunner Assertion Failure via Embeddings-Only Requests

Publication date: 2026-11-13

Last updated on: 2026-11-13

Assigner: VulnCheck

Description
vLLM from 0.28.0 before 0.30.0 contains a reachable assertion vulnerability in GPUModelRunner._init_xdrope_positions that allows remote attackers to crash the engine by submitting embeddings-only requests. Attackers can send a /v1/completions request containing only prompt_embeds to XD-RoPE models with --enable-prompt-embeds, terminating the EngineCore and API server for all clients.

CVSS Scores

EPSS Scores

Probability:
Percentile:

Meta Information

Published
2026-11-13
Last Modified
2026-11-13
Generated
2026-11-13
AI Q&A
2026-11-13
EPSS Evaluated
N/A
NVD
CVE-2026-126398
EUVD
EUVD-2026-125167

Affected Vendors & Products

Showing 1 associated CPE
Vendor Product Version / Range
vllm_project vllm From 0.28.0 (inc) to 0.30.0 (exc)

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

vLLM versions 0.28.0 to 0.29.x have a denial of service vulnerability when running XD-RoPE models with the --enable-prompt-embeds flag. An attacker can crash the entire vLLM engine and API server by sending a /v1/completions request containing only prompt embeds without any prompt text or token IDs. This occurs due to an unpatched assertion in the GPUModelRunner._init_xdrope_positions function that assumes prompt_token_ids will always be available.

Detection Guidance

To detect this vulnerability, check if your vLLM deployment uses XD-RoPE models with --enable-prompt-embeds enabled. Monitor API logs for /v1/completions requests containing only prompt_embeds without prompt text or token IDs. Look for crashes in the EngineCore or API server after such requests.

Impact Analysis

This vulnerability allows remote attackers to completely disrupt service by crashing the vLLM engine and API server. All clients relying on the service will experience downtime until the server is manually restarted. The attack requires no authentication if the API is exposed without an API key.

Mitigation Strategies

Upgrade vLLM to version 0.30.0 or later. Disable the --enable-prompt-embeds flag if not required. Restrict API access with authentication to prevent unauthenticated requests. Monitor for unusual requests containing only prompt embeds.

Chat Assistant

Ask questions about this CVE
Hi! I’m here to help you understand CVE-2026-126398. Ask me anything about the vulnerability, its impact, or mitigation strategies.
0/70

EPSS Chart

Security Radar

News mentioning this vulnerability
Loading related security news...