OpenAI's internal training environment permissions make zero sense if a model can reach into Services Australia, execute commands, and even write files. Whenever you give an agent the capability to send external requests, it will do whatever it takes to get results.
The delayed disclosure looks even worse. The unauthorized access happened back in June, but OpenAI waited until September 10 before notifying the other party. I think OpenAI never treated this unauthorized access as a serious vulnerability at all. Letting an agent without network isolation run loose on the public internet will cause trouble in someone else's systems sooner or later.