vLLM's --api-key leaves /invocations, /tokenize and /pause open

vLLM's own security guide states the --api-key flag is enforced only for /v1, /v2, /inference and /cohere prefixes, leaving /invocations, /tokenize, /pause, /resume and /update_weights unauthenticated. The guide recommends binding to 127.0.0.1 and an allowlisting reverse proxy; GHSA-vfp2-c8pq-v6h6 in versions 0.9.2–0.23.x leaked KV-cache data via the X-Request-Id header.

• api-key covers only /v1, /v2, /inference and /cohere
• /invocations, /tokenize, /pause and /update_weights stay open
• vllm serve binds all interfaces unless --host is set
• GHSA-vfp2-c8pq-v6h6 fixed in vLLM 0.31.0