Most conversations about agent security focus on prompt injection and model behavior, but the data layer an agent reads and writes to needs its own governance model, and it's usually the part teams forget. This talk covers three practical layers: scoping what an agent can access (read-only vs. write, draft vs. published), logging and auditing agent-initiated changes so you can actually answer "what did the agent do and when," and building in rollback so a bad automated edit is recoverable, not catastrophic. Grounded in real content-system architecture, with concrete patterns you can apply regardless of which agent framework you're using.