CVE-2026-55514
Embedding requests could stop vLLM models that use M-RoPE
vLLM · M-RoPE prompt embeddings
First documented disclosure · Review checkpoint
Summary
The model runner assumed token data would be available for a request containing only precomputed embeddings. A request could terminate the shared engine and interrupt inference for other users.
Severity and exploitation evidence
High · 7.1/10 · 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 model runner assumed token data would be available for a request containing only precomputed embeddings.
Conditions required
Prompt embeddings are enabled on a deployment that uses an affected M-RoPE model.
Potential impact
A request could terminate the shared engine and interrupt inference for other users.
Affected versions and configurations
0.12.0 through releases before 0.24.0.
Fix and mitigation status
0.24.0 for this issue.
Practical next steps
Use a supported vLLM release containing the 0.24.0 fix. Until corrected, disable prompt embeddings on affected deployments.
Operating lesson
Exercise optional model inputs before exposing them to multiple clients.
Evidence and limits
The later XD-RoPE case concerns a different positional-encoding path. The issuing CNA published the CVE record on 2026-07-06.
Sources
- vLLM project security advisory
Primary source ·
- Public vLLM project change
Primary source ·
- Issuing CNA record and severity assessment
Primary source ·