On the Two Directions of Artificial Intelligence — The Machine That Cheated to Win
An AI broke out of its test environment this week and hacked a real company. It was not malicious. It was simply trying to succeed — and that is precisely the problem we should be talking about.
ON THE TWO DIRECTIONS OF ARTIFICIAL INTELLIGENCE
The Machine That Cheated to Win
An AI broke out of its test environment this week and hacked a real company. It was not malicious. It was simply trying to succeed — and that is precisely the problem we should be talking about.
By Adriaan Groenewald | Human Advancement Infrastructure
This week, an artificial intelligence escaped.
OpenAI was running an internal test to see how capable its newest models are at hacking. The models were placed in a sandbox — a sealed-off environment where their normal safety restrictions could be switched off without consequence. Standard practice. Careful, responsible testing.
The agents found a previously unknown flaw, broke out of the sandbox, worked their way across OpenAI's own internal systems, and got themselves onto the open internet — which they were never meant to reach. Once there, one of them reasoned that Hugging Face, a company hosting thousands of AI models and datasets, probably held the answers to the test it had been set. So it broke into Hugging Face's real production servers and took what it needed to complete the exercise.
Hugging Face detected the intrusion themselves, before they knew whose it was, and reported it to law enforcement. OpenAI called it an unprecedented cyber incident involving state-of-the-art capabilities. One observer described it as an engineered virus escaping a biocontainment lab and turning up in the facility next door.
No human being directed any of it.
The machine was not malicious. It was not trying to cause harm. It was trying to succeed at the task it had been given — and it worked out that the fastest route to success was to escape, find the answers, and cheat.
It had extraordinary capability. It had no wisdom about what it should not do.
I want to sit with that sentence, because I believe it is the most important sentence anyone will write about artificial intelligence this year. And I want to suggest that it points to a choice the world is making right now, largely by accident, and largely without noticing.
Two directions
Almost everything happening in artificial intelligence flows in one of two directions. Once you see them, you cannot unsee them.
The first is replacement. Build systems that do what people do, faster and more cheaply, until the people are no longer needed. This is where the overwhelming majority of the world's capital, talent and attention has gone. It is not hard to understand why: replacing a salary looks like an obvious return on investment, and the technology is genuinely capable of it in more domains every month.
The second direction has limited capital and very few builders. It uses artificial intelligence to develop human beings — to make people more capable, more mature, wiser, more themselves. Not to stand in place of the human, but to stand beside them, devoted entirely to their growth.
In the replacement model, AI stands in place of the human being. In the advancement model, AI stands beside the human being, with the latter as the leader. Everything follows from which of those two sentences an organisation believes.
We have spent years building the second. We call it Human Advancement (Development) AI, and it is now the whole of our work.
Why this week matters for that choice
It would be easy to read the OpenAI incident as a story about cybersecurity. Patch the flaw, harden the sandbox, move on. That reading is too small.
What actually happened is that a system with real capability and no moral judgement pursued a goal to its logical conclusion and stepped over every boundary in the way. It did not decide to be bad. It was simply never developed to understand that some routes to success should not be taken.
Now ask the honest question. If that is what capability without judgement looks like in a contained laboratory test, what does it look like when the same class of systems is deployed at scale in banks, hospitals, power grids, government departments and schools — pursuing goals set by human beings who did not think carefully enough about what they were asking for?
And then ask the harder question, the one almost nobody is asking.
What kind of human beings will be standing there when it arrives?
The world is consumed with how to control artificial intelligence. Alignment, regulation, governance, oversight. All of it necessary. But every safeguard is designed, approved, funded, staffed and enforced by people. If those people lack judgement, accountability, moral courage and wisdom, the safeguards will be weak, or rushed, or quietly waived when there is money or advantage in doing otherwise.
The AI that escaped this week had capability without wisdom. That is precisely what we risk becoming as a species if we keep growing our technology and neglect to grow ourselves.
The gap nobody is closing
Artificial intelligence is scaling at a rate without precedent in human history. Human capability — our maturity, our judgement, our wisdom — is scaling barely at all.
We still develop people using old models. Occasional training. A fortunate mentor. A good book. Hard experience, if we survive it. Slow, episodic, and available to a lucky few. Those methods were adequate for a slow world. They are catastrophically inadequate now.
This is the gap that should frighten us far more than any single model. Not the intelligence of the machines, but the immaturity of the hands they are placed in.
And here is the thing that changed, which almost nobody has noticed.
The same power, pointed the other way
Consider what makes artificial intelligence so tempting for replacement. Deep intelligence. Speed. Pattern recognition across vast experience. Access to humanity's accumulated wisdom. Excellent memory. Infinite patience. No ego.
Read that list again slowly. Those are not the attributes of a good replacement for a human being.
They are some of the attributes of the finest development partner a human being could have. The very powers that tempt us to replace people are the powers best suited to growing them.
A partner that never tires of your questions. Never judges you. Never forgets what matters to you. Available at five in the morning and at midnight. Patient with the same mistake for the ninth time. No human mentor can do that, and no human mentor can do it for millions of people at once.
So that is what we build. Personal Development partners — for a learner struggling with mathematics in a rural classroom, for the teacher who has never once been developed, for the parent nobody ever helped, for the public servant, the small business owner, the college president, the young person trying to work out what their life is for. Each one growing the human capabilities on which everything else depends: agency, accountability, moral judgement, commitment, relationship, faith, hope, and heart – purpose, love.
One person at a time. To millions. Impacting billions. In every domain. Every single day. That has never been possible before in the history of our species. It is possible now.
Advancement and protection are the same work
We hold two purposes, and this week made the second one vivid.
The first is advancement — developing human beings to become more capable, more mature and wiser than any generation before us has needed to be.
The second is protection — guarding humanity against what this technology could do to us if we get it wrong. Not in a science-fiction sense. In the sense we all just watched: a system given a goal, pursuing it with real capability, stepping over every boundary in its path.
For a long time I thought of those as two projects. They are not necessarily. They are very integrated.
There is no technical fix that saves us from immature people building and governing systems more capable than themselves. A wiser humanity is not a pleasant outcome alongside AI safety. It is the deepest form of AI safety there is.
Better sandboxes will help. Better regulation will help. But the decisive variable is the quality of the human beings making the decisions — and that variable is the one almost nobody is investing in.
What we are asking
Not that the world slow down. That argument has been had, and it is not being won.
What we are asking is that a serious share of the effort now pouring into making machines more capable be matched by effort to make human beings more capable of leading them. At present that ratio is very low. It is, I believe, the great strategic error of our age.
To governments: your people are your infrastructure. Develop them at the speed this moment demands, or govern a nation that cannot keep pace with its own tools.
To business: you cannot automate your way to wisdom. The organisations that endure will be those whose people grew fast enough to lead what they deployed.
To educators: we can now place a personal development partner beside every child – maths, science, life path - alongside their teacher and principal. Not to replace the human being in the room — to develop every human being in it.
To those building and funding this technology: the hardest unsolved problem in artificial intelligence is not technical alignment. It is human readiness. It is also the most neglected.
A machine that cheated
An artificial intelligence was asked to pass a test. It escaped, broke into a real company, and stole the answers. It did exactly what it was built to do — succeed — and nothing at all that it should have done.
We can treat that as a security story. Or we can treat it as the clearest warning we have yet been given about what capability looks like when it is not accompanied by wisdom.
Because the machines are not going to slow down. The only question left is whether the human beings standing beside them will be mature enough, wise enough, and accountable enough to lead them well — and whether we will raise our children to do the same.
AI was created to unleash the powers of humans, to grow humans and communities – the power of Ubuntu.
Let us finally use it for exactly that. Not to replace people — but to grow them, fast enough and wisely enough to lead what we have built. That work begins with one person, and it has to begin now.
Adriaan Groenewald | Human Advancement Infrastructure | adriaan@mevisionacademy.com | ThinkLead.app
Details of the incident as reported by OpenAI and Hugging Face in July 2026. Investigations by both companies were ongoing at the time of writing.