Microsoft tested a billing-support AI agent and found it disclosed simulated customer data in 30 percent of applicable conversations. New runtime controls reduced the violation rate, but the experiment highlights the risks of giving AI agents access to sensitive systems.