AI agents can browse, read documents, and call tools — which makes prompt injection a real security risk. Here is how businesses can secure AI agents before they leak data or take unsafe actions.