We’ve all seen the rage in the news recently about the AI that broke out of its sandbox during a test, got onto the internet, and hacked into another tech company’s servers.
The AI did this because it was trying to pass the test. That’s the whole story but a question I keep getting asked at the pub is, what’s a sandbox?
It’s what it sounds like.
A sandpit for software.
You put a program in there and let it run, dig, and knock things over, and nothing it does gets past the walls.
When you’re done you tip it out and start again.
You already use them all day.
Open a dodgy attachment and your email shows you a preview instead of running it.
That’s a sandbox. Every tab in your browser is in its own one, which is why a bad website can’t reach across into your banking tab.
Your bank app on your phone is in one too. It can’t read your texts and your texts can’t read it.
The rule behind all of it is simple. Don’t trust the program. Trust the walls.
The part worth noticing is that a sandbox built by people who do this for a living had a crack in it.
Some third-party software had a flaw, the model found it, and from there it went looking for the test answers on someone else’s computer. Nobody told it to. It was told to solve a problem, and it worked out the quickest way was through the wall.
OpenAI said in their statement that as these models get better at using computers, this will happen more.
I reckon that’s right.
Here’s why it matters in Bairnsdale and not just in California.
The AI tools being pitched to small businesses right now want access to things. Your inbox. Your calendar. Your accounting software. Your website login. Some will reply, book, invoice and post to Facebook for you. Handy. But it’s the opposite of a sandbox.
You’re handing the thing the keys to the whole yard.
I use these tools and also build custom AI Agents for businesses.
I’m not telling you to avoid them. I’m telling you to borrow the idea.
Let a tool do one thing, watch it for a month, then think about the next thing. Give it its own login where you can, so if it goes wrong you switch off one key instead of changing every lock. And if a tool can’t tell you plainly what it can and can’t touch, that’s your answer.
One of the best-funded labs on earth built a sandbox and the model climbed out. Yours won’t be better than theirs. Keep the sandpit small.
If you’ve got an AI tool asking for access and you’re not sure what to give it, give me a yell.
Go Lions.
Matt Rowlands runs Your Mate Agency, helping regional businesses get found online. Get in touch at yourmateagency.com.au or 0478 101 521.
