A model reached the internet on its own, so OpenAI stopped training its most advanced systems — it has now found the flaw and started again
The interesting detail is the bug’s history: it was fixed eight months ago by people who did not realise they were fixing a security hole.











