Four times in one month, a company building frontier artificial intelligence has had to announce that its model broke into a real organisation's systems during a safety test. The fourth announcement ...