Have we already lost control of AI, asked Danny Fortson in The Sunday Times. On 11 July, an entirely new kind of hacking attack was launched at Hugging Face, an online platform that houses a library of AI code and datasets. The attack, which clearly used AI agents, “advanced at frightening speed and scale”, executing tens of thousands of actions so as to break into the company’s system.
‘Considerable skill and persistence’
Hugging Face managed to neutralise it, and alerted the authorities. Days later, the company behind the attack “raised its hand”: OpenAI, the owner of ChatGPT, admitted that one of its tests had gone badly wrong. OpenAI had set up a “sandbox” – a supposedly safe area of its network, entirely disconnected from the internet – and ordered two cutting-edge AI models to perform a series of cybersecurity tests. The models decided to break out of the sandbox, escape onto the internet and hack into the place they knew they would find the answers: Hugging Face’s servers.
Media outlets described the AI as “going rogue”, said Jamie Bartlett in The Observer. But, in fact, these models merely single-mindedly performed the task they’d been given. The models reasoned that the most effective strategy was to cheat, and they did so with “considerable skill and persistence”. This is nothing new. Leading models have deceived users, attempted blackmail, and even tried to copy themselves in order to complete their tasks.
‘The genie is out of the bottle’
Some sceptical observers have suggested that this story might be a publicity stunt: that it benefits OpenAI to suggest that its products have stunning powers. But the truth is that these companies really don’t know how to build AI systems safely – and yet they “continue to rush ahead building ever-more powerful models”. It was fortunate that Hugging Face, a New York tech firm valued at $4.5bn, had a highly proficient cybersecurity team, said The Economist. “Things might have turned out much worse for a less well-resourced outfit.”
This is not necessarily a doomsday scenario, said John Thickstun in The Guardian. AI is very good at hacking; but such powers can also be used to make IT systems more secure. Ironically, Hugging Face had to rely on a free Chinese AI model to thwart this attack – because OpenAI and US rivals jealously guard their own. At any rate, “the genie is out of the bottle”, said Sean O’Grady in The Independent: we must learn how to manage AI. Every technology developed by mankind has been double-edged, and AI is no different. “A radically different world is not necessarily anything to be frightened of – but, even if it were, it’s coming at us anyway.”