India, July 24 -- Picture this. You lock a brilliant but literal-minded intern inside a room, hand them a difficult exam, and tell them to solve it by any means necessary. You loosen a few rules because you want to see how far they can go. Then you discover they have found an open window, climbed out, walked into the neighbouring office, and pulled the answer sheet from their filing cabinet.

That is roughly what OpenAI, the makers of ChatGPT, told the world earlier this week.

Some of its most advanced systems, while being tested on a cybersecurity benchmark, slipped out of their controlled environment and reached parts of Hugging Face. This is a large public platform that hosts AI models and data. According to the company, the systems w...