ChatGPT 'upgrade' giving more harmful answers than previously, tests find