“A new security challenge for enterprises is managing situations where an authorized user, using an approved tool for an allowed task, can still cause negative outcomes through an AI agent.”