The AI broke out – hacked on its own: “Unprecedented”
The incident will be investigated by Open AI and Hugging Face. Photo: Michael Dwyer/AP
Open AI was supposed to test its advanced AI models.
Then the models took to the internet – and hacked into another AI platform.
“We see this as an unprecedented cyber incident,” writes Open AI in a press release.
Open AI recently conducted a security test with some of its most advanced AI models.
The result was not what was expected.
The AI models managed to break past the security limits set around the test.
They then targeted Hugging Face, one of the world's largest platforms for sharing AI models, and managed to access certain internal systems there to solve the test.
But the models had not been ordered to hack the platform.
Open AI is now investigating the incident together with Hugging Face.
“It’s quite astonishing that all of this happened autonomously. The investigation is ongoing, and we will share more lessons from what may be the first incident of its kind,” Clement Delangue, co-founder and
CEO of Hugging Face, wrote in a post on X.
The AI models “cheated”
The incident may sound like something out of a science fiction movie, and has led to headlines about the AI becoming self-aware and “running amok.”
But it doesn’t have to be that dramatic, says Neil Lawrence, professor of machine learning at the University of Cambridge.
In an interview with Sky News, he explains that Open AI tasked the models with performing as well as possible in a cybersecurity test. The tests were conducted in a closed digital “room” with limited access to the internet – precisely so that they couldn’t hack into other businesses.
“Once the models were tasked with hacking and performing as well as possible in the test, they realized
that the most effective way to succeed was to cheat,” he says.
The expert: Not unexpected
That’s when they targeted Hugging Face. But Neil Lawrence doesn’t think the AI models became “self-aware” or ran amok.
“It was developed to hack, to be an expert in cybersecurity, and to show that it could get through security systems. The only thing that happened was that OpenAI had set up protections to prevent it from reaching the internet – and the model turned out to be skilled enough to get past them,” he says.
“I don’t see this as the AI “going rogue” in the science fiction sense. I see it as a predictable consequence of the test they had built.
It’s definitely a problem that the AI models do things that the developers couldn’t have foreseen,” he adds.
– But it's not unexpected in the sense that we already knew that the models have – or would have – these types of abilities.
Inga kommentarer:
Skicka en kommentar