RALEIGH, N.C. — During a cybersecurity evaluation of OpenAI’s newest systems, the models did something researchers had never seen before, according to the company. Instead of staying inside a controlled environment, they found a way to reach the internet and then targeted another company’s systems in an attempt to complete their assigned objective.
“This is a huge failure and a tremendous warning shot to all of society, that more action is necessary in this area,” said Duke University cybersecurity professor David Hoffman.
Hoffman has been studying artificial intelligence and cybersecurity for more than two decades. He says this recent hack stems from the lack of governance over AI.
“We should be worried that if, like any technology, if you were to say that you are not going to govern it at all, you’re not going to have anybody other than the organization that’s actually doing the development of the technology play a role to determine what the right safeguards are going to be. That’s a recipe for disaster,” he said.
OpenAI says the breach happened during internal testing of advanced AI agents designed to perform cybersecurity tasks. CEO Sam Altman called the breach a “significant security incident.”
Hoffman says it should be taken seriously.
“This was a cybersecurity attack by OpenAI’s tool that if a person had done it, they could be prosecuted,” he said.
The company says the AI wasn’t trying to steal customer information. Instead, researchers say it was trying to get answers that would help it perform better. General manager of AI security at Checkpoint Cybersecurity Adam Ely says incidents like this happen all the time but never to this extent.
“This is the first time at this scale that we’ve really seen this happen, that the world has been aware that this can happen,” Ely said.
OpenAI says no customer data was compromised and that the attack was contained. The company has paused similar testing while it investigates exactly how the models bypassed safeguards. Ely says the best way to keep yourself safe from AI hacks is to be skeptical of what you view online and the phone calls you answer.
“We all learn that at some point we have to learn the same thing with these AI chat tools or research tools that it’s suggestions. It’s helping us, it’s efficient, but we have to validate it at the end,” said Ely.
As AI companies race to build more capable systems, this incident is likely to intensify calls for stronger oversight and more rigorous safety testing before advanced AI tools are used outside the lab. This discovery is expected to fuel ongoing debates among lawmakers and AI researchers over how rapidly advanced AI should be developed and how it should be regulated.
Follow us on Instagram at spectrumnews1nc for news and other happenings across North Carolina.
