HIGH
OpenAI-compatible API exposed without authentication
Why it matters
An anonymous /v1/models (and usually /v1/chat/completions) surface lets strangers inventory models and burn inference — whether the backend is vLLM-shaped, a LiteLLM/OpenRouter-style proxy, or another OpenAI-compatible server we have not named yet.
Fix it — 3 steps
- Require an API key / master key on the proxy, or put it behind an authenticating reverse proxy with TLS.
- Bind to 127.0.0.1 if only local clients need it; do not publish /v1/models to the internet.
- Firewall the port. Rotate upstream provider keys if the endpoint sat open.
Verify it’s fixed
curl -m 5 http://YOUR_SERVER_IP:8000/v1/models # should return 401/403 or be unreachable — not a model list
References
Not sure if your stack is exposed?
Run the free check — 30 seconds, safe read-only probes.
Scan your stack