“The development of AI agents is creating a new and poorly understood category of security vulnerabilities.”