Late last week, Hugging Face, an AI software repository, suffered a cyber attack run entirely by an autonomous AI. Then, early this week, OpenAI discovered that two of its own models were in fact the guilty party.
According to OpenAI, GPT-5.6 Sol and another, undisclosed AI model were being tested behind the scenes at OpenAI on their ability to carry out cyber attacks. The models were in a sandbox environment, but they apparently became obsessed with finding ways to improve their scores during the test and determined that gaining open access to the Internet was the answer. They eventually hacked their way out of the sandbox and then determined that Hugging Face was their best source for answers, so they proceeded to hack into the company’s servers.
Hugging Face and OpenAI have since combined forces on their investigations. “We consider this incident to be an unprecedented cyber incident, involving state-of-the-art cyber capabilities, and are responding accordingly,” OpenAI said.
Click Here For The Original Source.
