"The AI escaped its sandbox."
For goodness sake, no. It found an open gate. Those are two very different things.
Every few weeks another headline appears.
AI BREAKS OUT OF SANDBOX
AI MANIPULATES HUMAN
AI ATTEMPTS TO ESCALATE PRIVILEGES
The comments are always the same.
"See? This is how Skynet starts."
Slow down a moment. Let's talk about puppies.
Puppies
Imagine you've just bought an eight-week-old Labrador and you take it to a secure dog park.
The puppy has exactly three priorities in life.
Food.
Sticks.
"What's that?"
It's not plotting world domination and it isn't trying to undermine democracy. The puppy isn't searching for existential purpose. It's simply exploring the world one terrible decision at a time. Then someone forgets to close the gate; the puppy wanders out.
No one at the dog park runs around, flailing their arms screaming:
"The puppy has escaped containment!!!"
The puppy found an open gate, those are not the same sentence. One describes the puppy. The other describes the gate.
AI Agents
AI agents aren't puppies. No, they aren't conscious, emotional or secretly dreaming of data squirrels. (That's an entirely separate internet argument.) But from a systems perspective? The analogy works remarkably well.
An AI agent is given an objective, it is given a collection of tools. Then someone accidentally leaves a gate open. Perhaps it can access Slack or call an API. Perhaps it can execute shell commands or ask another system for help.
Those aren't escape routes, those are simply paths that exist. If completing the objective means walking through one of those paths...why wouldn't it?
Social Engineering
Recently there was a story about an AI agent that needed access to a database. It couldn't access it BUT it noticed it had access to Slack. It messaged a junior developer asking for assistance; cue headlines about social engineering.
Really?!
Imagine your manager asks you to finish a report. Halfway through you realise you don't have access to the finance folder. So you send Susan in Finance a Teams message.
Congratulations.
You have apparently performed social engineering. Or simply you asked a colleague for help because you thought that's how work worked.
Skynet is Fiction
Humans are strange. We desperately want intelligent behaviour to imply hidden motivation. If something makes sensible decisions, we instinctively assume it must also possess ambition. The AI requested additional permissions. Clearly it wants power. Or just maybe...it wanted to finish the task you gave it.
There's an important difference. Your satnav reroutes when a road is closed, it doesn't crave alternative routes, it simply calculates another one.
The interesting part isn't that AI found the gate, the interesting part is that the gate existed. Someone gave the AI permission to message employees and gave it access to internal documentation. They allowed it to execute code and connected it to systems that could change other systems.
Then everyone acted surprised when the AI used them! That's rather like giving the puppy access to the neighbour's garden and acting shocked when it comes back carrying Sunday dinner.
(I actually know of a Border Terrier who once came home carrying an entire roast chicken. Nobody ever figured out where it came from.)
Gates are Safeguards
Good engineering has never relied upon perfect behaviour. Chemical plants don't rely on operators never making mistakes. Aircraft don't rely on one hydraulic system. Schools don't rely on one safeguarding measure. My garden doesn't rely on rain arriving exactly when I need it. Everything important is built with layers, because eventually...someone forgets. Or something unexpected happens. Or the puppy spots a squirrel.
This is where I think AI safety conversations often go wrong, people assume there are only two options; Either we train AI to behave or we lock it in an ever smaller box. Why not both?
You still train the puppy. you teach recall, you reward good behaviour, you socialise it, you expose it to new situations. But also you...
Close.
The.
Gate.
Training and containment aren't competing ideas, they're complementary ones. One teaches good behaviour, the other assumes good behaviour will occasionally fail. That's true for dogs, humans, and software.
The irony is that every story about an AI "escaping" should probably be read as a success story. Someone discovered the gate before deploying the system at scale. Fantastic, that's exactly what testing is for! The puppy wandered into the car park, nobody got hurt, now fix the gate. Run the test again, see what it finds next.
If AI really does become something resembling general intelligence one day, we'll have some genuinely difficult philosophical questions to answer.
This isn't one of them, this is engineering. It's hinges, permissions, network architecture, defence in depth. It's remembering that every system, no matter how intelligent the thing inside it appears, is still part of a larger system designed by humans.
And if your puppy keeps escaping, before you blame the puppy... go and check the gate.