What happened
vLLM's chat-completions endpoint allows a remotely supplied chat_template value to drive uncontrolled resource consumption during Jinja rendering, a denial-of-service risk on shared serving infrastructure.
Why it matters
Consistent with the broader pattern of vLLM resource-exhaustion issues this cycle, this affects any multi-tenant OpenAI-compatible vLLM deployment that allows callers to influence the chat template, risking availability for all tenants on that instance.
Attack vector
A remote caller can manipulate the chat_template argument passed to /v1/chat/completions to cause excessive resource consumption during Jinja template rendering, degrading the shared inference service.
Affected systems
vllm-project vLLM, versions up to 0.27.1
Mitigation
Track and apply the upstream vLLM fix; in the interim, restrict who can set custom chat_template values on multi-tenant endpoints and apply request-level resource limits.