Securing the AI Inference Stack: GPU Memory Isolation, Model Serving Hardening, and Self-Hosted LLM Infrastructure Security
Self-hosted LLMs introduce attack surfaces below the application layer: GPU VRAM residuals between tenants, unauthenticated serving APIs, and unverified model weights on disk. This post maps the infrastructure threat model and provides a hardening checklist.