Operators Harden Kubernetes LLM Deployments With Policy

On March 30, 2026, Nigel Douglas (CloudSmith) warns that Kubernetes alone cannot enforce LLM-specific security controls and outlines four operational risks—prompt injection, sensitive-data leakage, supply-chain compromise, and excessive agent agency—that operators must mitigate when running models in-cluster. He advises implementing a separate policy layer (an LLM-aware gateway) rather than embedding checks in runtimes like Ollama, and lists tools such as LiteLLM, Kong AI Gateway, Portkey, and kgateway.