Jack Tame: AI is breaking out

Author: Newstalk ZB,
Publish Date: Sat, 25 Jul 2026, 11:12am

Follow the podcast on

It’s been yet another week of yet more unsettling news. The conflict in the Middle East appears to be escalating again. As of this morning, we face even higher tariffs in the U.S than under President Trump’s previous measure. But if there’s one piece of unsettling news I think we probably haven’t grappled with sufficiently, it’s the latest humbling episode with one of the world’s biggest AI companies.

OpenAI admitted yesterday that one of its new models has autonomously broken free of its shackles, escaped from the walled-off environment where it was supposed to be safely tested, gained access to the wider internet, and successfully hacked another company.

Legally it's a grey area. There's a pretty strong argument that this was a crime. If a person did it, theoretically they could be charged. But it wasn’t a person. It was an autonomous AI model going to fairly extreme lengths in order to solve a problem that had been posed by its creators. Again, it was supposed to be contained. Those same masters who posed it a challenge had thought it was confined in what they call a sandbox, a sort of digital playpen where you can wall in a model and stop it from escaping to the outside world.

But like a toddler who pulls themself up and climbs out of a cot, this model got free. AI models are now so good at exploiting cyber vulnerabilities, they can exploit the vulnerabilities of their creators. Ultimately, the model solved the problem it had been set. But at some cost.

There were heaps of details that made this whole episode particularly interesting. For starters, it looks like OpenAI didn’t really understand what its own model was doing. They couldn’t tell in real time that it had escaped and was roaming the internet at large. They couldn’t pull a metaphorical power cable out of the wall, because they only realised what it had done after the fact.

But if there’s one thing to be grateful for in this instance, it’s the target of the AI hack. It wasn’t a bank or a crypto-currency exchange or a weapons system. It was a start up called Hugging Face which works as an interface for open source AI. If anything, the boss of Hugging Face sounded kind of stoked that his business had been targeted.

Of course this is the sort of thing that experts have been warning about for ages. At the most extreme end of the AI doomerism scale, they’ve suggested these systems could ultimately decide humans are an inconvenience and that the world would be a much better place if computer systems were in charge. And arguably one very important step on the road to AI Armageddon is this - a model breaking free.

If you can’t be bothered engaging with AI, or even if you think AI is overhyped...this week marks a significant milestone. In one sense, it is incredibly impressive that an autonomous system, operating completely free of human control, is able to pull off something with this complexity. In another sense, it is incredibly unsettling. If this is where the technology is today, who can honestly say where it might be in five or ten years’ time?

LISTEN ABOVE