AI Stack Alert Medium CVE-2026-115911 For AI Eng, AppSec, SOC / IR, CISO AI-assisted
vLLM <0.32.0: chat_template rendering DoS wedges all chat endpoints (CVE-2026-115911)
vLLM renders chat templates in a Jinja2 ImmutableSandboxedEnvironment with no wall-clock timeout, output-size cap, or loop/binop limit. The render runs in a thread pool with max_workers defaulting to 1, and a single crafted short template can block it until the process restarts. Affected versions are vllm >= 0.19.0 and < 0.32.0; the fix is in 0.32.0.
Why it matters
An authenticated API client, or a malicious model repo loaded with trust_remote_code=False, can deny chat inference to every tenant of the affected API-server replica. The string-multiplication variant can OOM-kill the API-server process.
Do this
- Upgrade vLLM to 0.32.0 or later on all API-server replicas.
- Until patched, restrict which clients can reach /v1/chat/completions, /v1/responses, /tokenize, /pooling and /invocations at the gateway, and limit API keys to trusted tenants.
- Review chat templates in tokenizer_config.json and any --chat-template input before loading untrusted model repos.
- Alert on chat endpoint timeouts, API-server process restarts, and OOM kills on vLLM hosts.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.31%
Signals
- NoCISA KEV
- 6.5CVSS
- 0.31%EPSS
- 27/100Priority
- 85%Confidence
Evidence
“The render runs in a `ThreadPoolExecutor` whose `max_workers` defaults to **1**, and CPython worker threads cannot be interrupted.”
“The string-multiplication variant OOM-kills the process. In a multi-tenant deployment this is a cross-tenant denial-of-service from any one tenant.”
“a model repo whose `tokenizer_config.json` carries the hostile template wedges every replica that loads it on the first legitimate chat request, with `trust_remote_code=False`.”
Sources
Affected / related
vLLM, chat_template Jinja2 rendering
Published 2026-10-28 21:14 UTC · updated 2026-11-08 02:47 UTC · AI-assisted summary, checked against quoted sources · How we score