New research shows how attackers can use security alerts and blocked events to manipulate and hijack AI agents.