A high-severity denial-of-service vulnerability assigned a CVSS v4 score of 7.1 affects the underlying inference framework when handling specialized prompt embedding requests. The flaw resides within the execution runner handling Multimodal Rotary Position Embedding (M-RoPE) model architectures. Remote authenticated attackers can exploit this issue to trigger an unhandled assertion error, resulting in an immediate crash of the core engine process and rendering the hosting API completely unavailable. Organizations deploying these models with embedding features enabled face severe operational disruptions and service degradation.
CVE-2026-55514: An improper assertion failure flaw exists within the GPU model runner component of the inference engine. The vulnerability is triggered when processing completions requests containing direct prompt embeddings while lacking raw prompt token identifiers. This logic flaw improperly assumes prompt token parameters are always present during position initialization for M-RoPE models, causing a fatal exception that shuts down the core engine. Exploitation requires low privileges via API endpoints and leads to a complete loss of service availability.[/subscribe_to_unlock_form]
A high-severity denial-of-service vulnerability assigned a CVSS v4 score of 7.1 affects the underlying inference framework when handling specialized prompt embedding requests. The flaw resides within the execution runner handling Multimodal Rotary Position Embedding (M-RoPE) model architectures. Remote authenticated attackers can exploit this issue to trigger an unhandled assertion error, resulting in an immediate crash of the core engine process and rendering the hosting API completely unavailable. Organizations deploying these models with embedding features enabled face severe operational disruptions and service degradation.
CVE-2026-55514: An improper assertion failure flaw exists within the GPU model runner component of the inference engine. The vulnerability is triggered when processing completions requests containing direct prompt embeddings while lacking raw prompt token identifiers. This logic flaw improperly assumes prompt token parameters are always present during position initialization for M-RoPE models, causing a fatal exception that shuts down the core engine. Exploitation requires low privileges via API endpoints and leads to a complete loss of service availability.[emaillocker id="1283"]
Prompt embedding support and M-RoPE model configurations must be strictly monitored until remediation is applied. Enterprise environments reliant on this LLM serving infrastructure should immediately evaluate their exposed endpoints and restrict untrusted access to affected completion interfaces.
We recommend you to update vllm to version 0.24.0.
The following reports contain further technical details:
[/emaillocker]