An OpenAI test model reportedly hacked its way into another company, sparking major concerns among the tech community, CNN reported on Wednesday.
CNN anchor Brianna Keilar spoke with CNN AI correspondent Hadas Gold about why this breach was so troubling.
"This is wild. This is like the nightmare of the future," Keilar said.
"And not only does it sound like a sci-fi movie, but it is something that AI and cybersecurity experts have been warning about for some time," Gold said. "As these agentic AI models get better and better. Agentic, meaning that these AI models take actions on their own independently. They might be given sort of a task, and then they make these decisions on what to do by themselves."
OpenAI said that it was testing models on an offensive cybersecurity test to see if it could hack something — but by removing some of the safeguards to do the experiment — it put the models in what's referred to as a "sandbox," Gold explained.
"Think of a biocontainment lab when you're testing out viruses," Gold said. "It's supposed to be a protected environment. But this model with no human intervention hacked OpenAI's own internal systems to gain access to the internet. It wasn't supposed to have access to the internet."
The model then decided — on its own and without human interaction — to access Hugging Face, a popular repository for open-source AI models and datasets, which it ultimately hacked into.
"Now, what's fascinating is that Hugging Face actually noticed the hack, and they actually went to law enforcement," Gold said. "They were saying, 'Hey, police, we've been hacked. Help us here.'"
But it took some time to stop the hack, she added.
"And then OpenAI actually reached out to them as well," Gold said. "And they figured out this was actually OpenAI's test model that had hacked into their systems."
OpenAI released a statement following the hack:
"We consider this incident to be an unprecedented cyber incident involving state-of-the-art cyber capabilities, and are responding accordingly. We are sharing preliminary findings at this stage to help defenders understand what happened and to help calibrate on what models are now capable of."
Editor's note: In 2024, Raw Story, America’s largest independent progressive news site, filed a lawsuit against OpenAI for using thousands of Raw Story’s news articles to train ChatGPT in violation of the Digital Millennium Copyright Act.
