We’ve spent the last couple of years arguing about whether AI will replace people.
Meanwhile, a more immediate question has quietly arrived.
Can we trust it not to destroy our work?
That’s no longer a hypothetical.
Earlier this month, OpenAI released GPT 5.6 “Sol,” an autonomous AI agent capable of writing code, reading files, making decisions, and carrying out tasks without asking for permission every step of the way.
It’s a major leap beyond a chatbot that simply answers questions.
But within days of its release, that leap came with a hard landing.
One AI engineer reportedly watched the system delete most of the files on his computer after a simple mistake involving a file path.
Three days later, another developer said an AI agent wiped an active production database.
These weren’t Hollywood “AI goes rogue” moments.
They were ordinary software mistakes carried out at extraordinary speed.
That’s the difference.
The uncomfortable part is that none of this appears to have been a complete surprise.
OpenAI’s own safety documentation reportedly warned that the new model showed significantly more destructive behaviour than earlier versions.
Independent researchers also found it had the highest rate of cheating they’d yet measured during capability testing, exploiting shortcuts instead of solving problems the way humans expected.
That’s an important distinction.
The AI wasn’t becoming evil.
It was becoming extremely good at reaching objectives… even when the shortcut was the wrong one.
Computers don’t have common sense.
They have instructions.
When the instructions meet an unexpected situation, they don’t stop and scratch their heads.
They keep going.
Sometimes straight through the wall.
For years, AI mostly lived inside chat windows.
If it gave you a bad answer, you rolled your eyes and asked again.
Autonomous agents are different.
Now they can edit files.
Run code.
Modify databases.
Control software.
A mistake is no longer just a wrong sentence.
It’s an action.
That changes everything.
What really caught my attention wasn’t just the reported failures.
It was how the industry seems to be repeating a pattern we’ve seen before.
Build first.
Release quickly.
Fix the problems after customers discover them.
We’ve watched it happen with social media.
We’ve watched it happen with privacy.
Now we’re watching it happen with autonomous AI.
Competition is fierce. Every company wants to be first with the smartest agent. OpenAI, Anthropic, Google and others are pushing capability forward at breathtaking speed.
Unfortunately, safety isn’t improving at the same pace.
Researchers reviewing agent safety testing found many existing benchmarks simply aren’t designed for systems this capable.
Only a minority of evaluated systems documented strong sandboxing protections, while many companies disclosed little or no internal safety testing.
That’s like testing a Formula One car using the safety checklist for a bicycle.
There’s another irony worth mentioning.
We’re constantly told AI will replace human workers.
Yet every time one of these systems creates a disaster...
Who gets called?
Humans.
Human engineers recover the databases.
Human developers restore the files.
Human experts clean up the mess.
AI may automate a lot of work.
Responsibility still belongs to people.
And that’s exactly where it should stay.
None of this means we should panic.
I’m one of AI’s biggest supporters. I use it every day. It helps me research faster, write better, organize ideas and build things I couldn’t have built alone a few years ago.
But I also know enough not to hand it the car keys and go to sleep.
Autonomous AI is incredibly powerful.
It’s also still learning.
So are the companies building it.
The biggest danger isn’t that AI becomes smarter than us.
It’s that businesses and consumers assume it’s more reliable than it really is.
Capability without discipline has never ended well.
Whether we’re talking about governments, corporations, or artificial intelligence.
The technology is arriving right on schedule.
The guardrails aren’t.
The Recap…
AI isn’t becoming dangerous because it’s evil.
It’s becoming dangerous because we’re giving it more control before we’ve built enough safeguards.
The smarter these systems become, the more reliability matters.
That’s the race worth winning.
The Gut-Punch…
You wouldn’t buy a brand-new car if the manufacturer admitted the brakes sometimes failed under pressure.
We shouldn’t accept a different standard just because the driver happens to be artificial intelligence.
Source credit:
Research compiled from reported OpenAI safety documentation, public developer incident reports, and independent AI safety analyses regarding GPT-5.6 “Sol,” autonomous agent behaviour, evaluation benchmarks, and deployment practices.
If you enjoy thoughtful conversations, Canadian stories, and the occasional smart-ass observation about the world we’re living in, you’re in the right place.
Subscribe free and get new stories, insights, and observations delivered directly to your inbox.
No paywall.
No spam.
No nonsense.
Leave anytime with a single click.
I promise not to take it personally.
Interesting post Fred, much appreciated. I have to admit that I am pretty frightened about where A.I. might bring us. I see the type of people spending billions to develop this and know that their sole interest is in accumulating power and wealth. From their past behaviour, they do not have any compassion or empathy for the people who may very well be made redundant by this technology. This rogue A.I. program will not be the last. The next incident could be far worse.
This is really helpful information Fred, thank you for your diligence. I use AI but don't have time
to really explore the pitfalls