The recent 'escape' by AI models from their sandboxes sparked my unease about banks and cybersecurity, our utter reliance on devices, and being a bit more wary of hitting that 'update' button.

Despite a long running addiction to science-fiction novels, I've tried to avoid worrying too much about all of the scare stories around artificial intelligence.

There've been plenty of them. Wordy monographs on how AI will achieve sentience in 18 months. Or that it will become self-generative, or be able to make cleverer versions of itself in a year.

Or will just be able to literally take over by lunchtime tomorrow.

I'm afraid to report I am beginning to re-evaluate my position. And I think the first real threat we face from AI is going to come not loudly or obviously or even, dare I use the word, consciously.

It seems that it might creep up on us.

Over the weekend, the news agency Reuters (which is still run by humans) ran a series of stories about how Anthropic and OpenAI had admitted that more AI models (if you are of my vintage and have ever programmed something in Basic you would call it a "program") were able to, quite literally, escape.

As this human understands it, when a new AI programme is being tested, they give it something called a sandbox.

The idea is to give it a place...