Running AI in production introduces new infrastructure and security problems. This section covers isolation, permissions, model serving, deployment, observability, and the risks that appear when models can take actions.