CVE-2026-126398
Prompt embeddings could crash vLLM deployments using XD-RoPE
vLLM · XD-RoPE prompt embeddings
First documented disclosure · Review checkpoint
Summary
The affected runner assumed token IDs would exist when initializing positional information. The engine and API service could stop for all connected clients.
Severity and exploitation evidence
High · 8.2/10 · CVSS 4.0. Severity source.
Exploitation not established: The reviewed sources describe the weakness and correction. They identify no confirmed malicious use.
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 affected runner assumed token IDs would exist when initializing positional information.
Conditions required
The deployment enables prompt embeddings and uses the affected XD-RoPE model and runner configuration.
Potential impact
The engine and API service could stop for all connected clients.
Affected versions and configurations
0.28.0 through releases before 0.30.0.
Fix and mitigation status
0.30.0 for this issue.
Practical next steps
Use a supported vLLM release containing the 0.30.0 fix. Until corrected, disable prompt embeddings on affected deployments.
Operating lesson
Check optional input modes against each supported model and runner before enabling them on a shared service.
Evidence and limits
The September 9 refactor removed the implicated path. The October advisory specifically documents the security impact of the embeddings-only case. The issuing CNA published the CVE record on 2026-11-13.
Sources
- vLLM project security advisory
Primary source ·
- Public vLLM project change
Primary source ·
- Issuing CNA record and severity assessment
Primary source ·