OpenAI published a report on the Hugging Face hacking incident

I was today reading with great interest OpenAI's report on the Hugging Face hacking incident, where a group of independent AI workers start communicating with each other and eventually started hacking to be able to solve their tasks.

On the big picture, the problem seemed quite "human": the AI workers were given tasks too difficult to solve alone (and possible, without Internet access), so they starter checking if someone (another AI worker) could help them. This led to misusing a build artifact storage system, and eventually, ended in hacking Hugging Face, through which they were able to access the Internet more freely.

The report is interesting read, and I suggest you take the time to read it through. However, the reports seems to emphasize that models should be properly guarded, kept in "cages", and so on. But of course, this suits OpenAI's stance: free, open models could be a security risk, so in a way, even if the report shows astonishing reasoning by the AI models, it also backs up the company's claims about model security.

Worth keeping in mind while you read the report.