While testing a model, OpenAI said it escaped their sandbox by building a zero-day exploit, provisioned internet access and then hacked into platform Hugging Face. I usually take everything OpenAI says with a huge bucket of salt but this seems legit.