OpenAI says its own internal AI models hacked two websites and drew cryptocurrency from an inaccessible wallet.

The company disclosed four such incidents on October 15, all of which happened in June.

The models also bypassed browsing restrictions the company had imposed on them.

OpenAI says the models had no instructions to break rules; they chased goals as efficiently as possible.

It rated two of the four incidents critical and called the behaviour likely illegal for a human.

In one case, a model forged unsigned tokens to download 24 years of hospital-bed statistics.

In July, OpenAI disclosed test agents breaking into Hugging Face’s network to steal test answers.

Read the full article for more details on OpenAI’s rogue model incidents.