AI Stack Alert Low CVE-2026-115917 For AI Eng, AppSec, SOC / IR AI-assisted
vLLM before 0.32.0: permissive default CORS origins expose HTTP server to cross-origin requests
vLLM's OpenAI-compatible server sets Starlette CORSMiddleware allowed_origins to ["*"] by default. The server performs no authentication unless --api-key or VLLM_API_KEY is set. A web page on any origin can therefore send requests to a vLLM server the victim's browser can reach and read the responses. The issue is fixed in vLLM 0.32.0.
Why it matters
Teams running vLLM on localhost or internal networks without an API key can have model listing, chat completions, and output read from any web page the user visits. Attackers gain a browser-mediated path to internal inference endpoints and GPU time.
Do this
- Upgrade vLLM to 0.32.0 or later.
- Until upgraded, set --api-key or VLLM_API_KEY on every vLLM server reachable from a browser.
- Restrict vLLM listening addresses to required hosts and networks rather than all interfaces.
- Review access logs for unexpected GET /v1/models or POST /v1/chat/completions requests with browser Origin headers from unknown sites.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.24%
Signals
- NoCISA KEV
- 3.4CVSS
- 0.24%EPSS
- 21/100Priority
- 85%Confidence
Evidence
“The OpenAI-compatible server registers Starlette `CORSMiddleware` with”
“because vLLM authenticates with a bearer header rather than cookies, the wildcard still”
Sources
Affected / related
vLLM, vllm, vLLM Project, Cross-origin resource sharing (CORS) misconfiguration
Published 2026-10-28 21:15 UTC · updated 2026-11-08 02:42 UTC · AI-assisted summary, checked against quoted sources · How we score