Security-oriented three-axis observability for sandboxed AI agents: model intent, app co…