[BUG] No authentication or authorization on any API endpoint — full control exposed #19

New issue

Closed

opened 2026-05-08 23:46:36 +00:00 by Mark · 0 comments

Mark commented

2026-05-08 23:46:36 +00:00

Owner

Summary

The vllm_mm backend has zero authentication or authorization on any endpoint. The CORS middleware allows all origins (*). Any client that can reach the API host can:

List, create, update, and delete model configurations
Start and stop model inference processes
Upload chat template files (arbitrary file writes)
Access GPU and system metrics
Proxy inference requests through running models

Affected

All endpoints in backend/app/main.py and all routers.

Risk Level

Internal/Dev deployment: The docker-compose uses network_mode: host which exposes ports directly on the host. If the host is reachable on a network, the API is fully open.
Production: Not suitable for any multi-tenant or internet-facing deployment without a reverse proxy handling auth.

Recommended Mitigation

Add an API key check middleware (simple X-API-Key header check)
Document that for production use, a reverse proxy (nginx, Caddy) should handle authentication
At minimum, add a configuration option to require authentication

Issue #1 (already open) about frontend UX — this is the backend security gap.

## Summary The vllm_mm backend has **zero authentication or authorization** on any endpoint. The CORS middleware allows all origins (`*`). Any client that can reach the API host can: - List, create, update, and delete model configurations - Start and stop model inference processes - Upload chat template files (arbitrary file writes) - Access GPU and system metrics - Proxy inference requests through running models ## Affected All endpoints in `backend/app/main.py` and all routers. ## Risk Level - **Internal/Dev deployment**: The docker-compose uses `network_mode: host` which exposes ports directly on the host. If the host is reachable on a network, the API is fully open. - **Production**: Not suitable for any multi-tenant or internet-facing deployment without a reverse proxy handling auth. ## Recommended Mitigation 1. Add an API key check middleware (simple `X-API-Key` header check) 2. Document that for production use, a reverse proxy (nginx, Caddy) should handle authentication 3. At minimum, add a configuration option to require authentication ## Related Issue #1 (already open) about frontend UX — this is the backend security gap.