gpu-hosted-asset-hardening

Hardening playbook for any long-lived GPU box running ML daemons (ComfyUI, Ollama, vLLM, Triton, Ray, JupyterHub, Hugging Face TGI). Covers threat model, network hardening, auth + quotas, capability sandboxing, secrets and storage, observability, patching cadence. NOT for one-shot serverless inference (use media-gen-deployment), NOT for incident response on a suspected-compromised box (use comfyui-incident-response), NOT for ecosystem-side supply-chain hygiene (use supply-chain-defense-for-ml-extensions).