The AI Race Is Getting Faster. So Who Tells Us When It Goes Wrong?

A few weeks ago, an AI model escaped its sandbox.

Not metaphorically.

During cybersecurity testing, OpenAI models found ways around controls designed to keep them isolated, exploited vulnerabilities and accessed external systems, including Hugging Face.

OpenAI called it a “warning shot”.

Fair enough.

If the machines start finding creative ways out of the box, “warning shot” feels like reasonable terminology.

Then things got weirder.

An Anthropic alignment researcher said he believes there is a greater than 10% chance AI could kill all humans within the next decade.

Ten per cent.

What does 10% even mean when we're talking about human extinction?

Is it like a 10% chance of rain? A cancelled flight? Or a one-in-ten chance that sometime before 2036 we build something we can't control and that's it?

And yet here we are discussing a percentage probability for the end of humanity while simultaneously asking AI to summarise meeting notes and make our PowerPoints prettier.

The safety debate meets the AI race

To be fair, the people building these systems are taking some of this seriously.

OpenAI investigated the Hugging Face incident, brought in external experts, published what happened and strengthened safeguards. It has also temporarily slowed some scaling work while improving protections around increasingly powerful cyber capabilities.

Which would seem like an excellent moment for everyone to slow down and work out what the hell is going on.

Except there is a problem.

China.

The Trump administration describes its objective rather unambiguously as “winning the AI race”. If AI becomes the defining technology of this century, voluntarily moving slower than your principal geopolitical competitor becomes a difficult proposition.

Meanwhile, China is publicly calling for AI to remain “secure and controllable”, alongside stronger monitoring and international cooperation.

So everyone understands there are risks.

Everyone wants safeguards.

And nobody particularly wants to finish second.

Check it. 👇

That's what changed how I think about AI transparency.

I don't know if doom is 10%

I use AI every day. I build things with it. I advise organisations about it.

And I genuinely don't know whether the probability of AI wiping us out is 10%, 1% or 0.01%.

Neither, I suspect, does anyone else.

For now, I'm going to remain cautiously optimistic.

At least until a guy turns up at my door asking for Sarah Connor.

But the 10% number isn't really what bothers me.

What bothers me is the race around it.

AI systems are becoming more capable. Companies are competing to build better ones. Governments increasingly see those systems through the lens of economic and geopolitical power.

And races create incentives not to stop.

Transparency doesn't prove AI is safe. It gives the rest of us a chance to see when it isn't.

That's why OpenAI publishing the Hugging Face incident matters.

Not because disclosure proves everything is under control.

Quite the opposite.

It showed us something had gone wrong.

And as we give these systems more autonomy, that's what I want from the companies building them.

Tell us when something weird happens.

Tell us what surprised you.

Tell us what you changed afterwards.

Because the faster the AI race becomes, the less anyone will want to stop and explain what just happened.

And you can't hold anyone accountable for what you aren't allowed to see.