Testing agent security policies against adversarial inputs — NullRun