AI security: Identity and access control

In our first 3 articles, we framed AI security as protecting the system, not just the model, across confidentiality, integrity, and availability, and we showed why the traditional secure development lifecycle (SDLC) discipline still applies to modern AI deployments. We also focused on guardrails and different architectural approaches such as dual LLMs and CaMeL to help protect against prompt injection and unsafe actions.This article completes the defense strategy by focusing on the backbone that makes guardrails enforceable in production—identity, authentication, authorization, and zero trus