AI Stack Alert Medium CVE-2026-115918 For AI Eng, AppSec, SOC / IR, CISO AI-assisted
vLLM Responses API bypasses trust_request_chat_template (CVE-2026-115918)
The vLLM /v1/responses endpoint accepts request-level chat_template_kwargs and does not apply the --trust-request-chat-template gate used by Chat Completions, render, tokenize, and pooling paths. A caller can supply a request-controlled Jinja chat_template even when the trust flag is disabled. Affected versions are vllm >= 0.16.0 and < 0.32.0; the fix is in 0.32.0.
Why it matters
Any authenticated caller of /v1/responses can submit attacker-controlled Jinja templates despite default-deny settings. Prior advisories tie request-controlled templates to CPU and memory exhaustion, so exposed vLLM servers face a denial-of-service risk.
Do this
- Upgrade vLLM to 0.32.0 or later.
- Until upgraded, restrict access to /v1/responses to trusted clients at the gateway or reverse proxy.
- Confirm that --trust-request-chat-template is not enabled on servers that do not need it, and that no request-level chat template is expected.
- Review access logs for /v1/responses requests containing chat_template_kwargs with a chat_template key.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.35%
Signals
- NoCISA KEV
- 6.5CVSS
- 0.35%EPSS
- 27/100Priority
- 80%Confidence
Evidence
“does not apply the `--trust-request-chat-template` gate that protects Chat Completions, render, tokenize, and pooling paths.”
“An authenticated caller who can submit `/v1/responses` requests can supply a request-controlled Jinja chat template despite the trust flag being disabled.”
Sources
Affected / related
vLLM, vllm, chat_template_kwargs
Published 2026-10-28 21:14 UTC · updated 2026-11-08 02:42 UTC · AI-assisted summary, checked against quoted sources · How we score