GPU VulnDB

Database/AI/ML frameworks & serving

vLLM: request-supplied chat template bypasses trust_request_chat_template and exhausts the server

CVSS 6.5CVE-2026-115918AI/ML frameworks & servingcurated

Impact

When no server-side chat template is configured, an authenticated caller can put an arbitrary Jinja template in chat_template_kwargs[chat_template] on /v1/responses. ChatParams.get_apply_chat_template_kwargs() keeps that value without enforcing trust_request_chat_template, so the operator's intended trust policy does not apply and the rendering can burn CPU and memory until the serving process is starved. On a GPU node this denies service to every tenant sharing that vLLM instance while the GPUs sit idle behind a stuck frontend. Affects 0.16.0 through 0.31.x.

Who can reach it

Any caller authenticated to the vLLM API that can reach /v1/responses on an instance started without a chat template.

What to do

Upgrade to vLLM 0.32.0 and restart the serving processes. Until then, configure an explicit server-side chat template and/or reject chat_template_kwargs at the gateway.

References

Related entries

All AI/ML frameworks & serving entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.