Since the huge leaps by the likes of OpenAI and Anthropic in the past few years, we have been swamped with comparisons to the Terminator movies and warnings from some corners that artificial intelligence will, ultimately, destroy life as we know it.
After all, Elon Musk – who admittedly has a penchant for hyperbole – has said there is at least a chance that “AI will kill us all”.
In April, Anthropic said it would limit access to its latest model given its potential for hacking, and there have already been plenty of examples where hackers have used an artificial intelligence model to boost their ability to beat companies’ cyber defences.
Given that background, what should we make of OpenAI’s revelation on Wednesday that one of its AI agents – essentially a program that doesn’t need human direction – broke out of a testing environment and into a start-up named Hugging Face by itself?
READ MORE
“The model chained together multiple attack vectors, including using stolen credentials and zero-day vulnerabilities to find a remote code execution path on the Hugging Face servers,” said OpenAI in a blog post.
In other words, it stole the equivalent of usernames and passwords and exploited a flaw in Hugging Face’s defences to gain access.
This is the sort of thing that was previously reserved for science fiction and dystopian novels but at this point is probably not that surprising. Something like this has been coming for some time.
It is easy to say that OpenAI – and other such firms – need to build stronger sandboxes to contain their programs, and as a society we need to rein in artificial intelligence. Much as Cantillon would like this to happen, that ship appears to have sailed.
Will the State’s plan for the Carlton Cinema site revitalise O’Connell Street?
This will act as an advert for OpenAI’s agents in some quarters. There are people who will take this as evidence of the great potential of AI and the need to keep pushing development.
For many others, this will underline the dangers of advancing AI too quickly.
Even if this isn’t the equivalent of Arnold Schwarzenegger in a leather jacket wielding a shotgun, this definitely won’t be the last time an AI agent breaks out.
