Jinja Rendering Vulnerability in vLLM by vLLM Project

CVE-2026-115918

What is CVE-2026-115918?

The vLLM inference and serving engine for large language models is susceptible to a Jinja rendering vulnerability between versions 0.16.0 and 0.32.0. The flaw exists in the /v1/responses endpoint, where an authenticated user can insert an attacker-controlled template into the chat_template_kwargs[chat_template] if no server template is configured. This condition allows the request to circumvent established trust policies, potentially resulting in CPU or memory exhaustion due to intensive Jinja rendering operations. The issue has been addressed in version 0.32.0.

References