Any AI that you don’t run locally is basically one big data breach.
How easily we leak confidential data via AI was made clear to me this week by a new employee. He had thrown all of a customer’s exploration data into his own ChatGPT account to make a nice schedule out of it. “Yes, but they don’t like it, because I took the name off.” Right, you still had the fixed ip’s on it, so a bit of self-respecting piece of intelligence doesn’t need a name to figure out who it’s about.
Painful but fortunately limited to a small set of IP addresses.
In the meantime, we have all become addicted to our daily ChatGPT, Claude or Gemini to get our tasks, life questions and relaxation done. Most individuals and SMEs run this entirely in public models, enterprise still almost half on other people’s, mostly US AI farmers.
Especially if you do this with your own account, you pay, as with so many other free internet services, with the data and insights you push into the model. In other words: you are the product and you pay with your own data.
And that also applies to the so-called walled gardens, the shielded, more expensive LLMs and models: you feed your data clearly legible to large, mainly American, third parties who then promise on their communion soul that they will not use it.
Foolish, because American law can force them to hand over the data, no matter what they promise. And agents and integration with the entire operating system and the applications that run on it open the door wide for the data gobblers to record almost anything.
We now know that the data from private accounts is also used outside your own sessions to train the model. I think this mass of data is the big engine under the hood for the enormous growth in capabilities of all players.
ChatGPT and Gemini do this by default, and Anthropic has also been training on chats from Claude Free, Pro, Max, and Claude Code since the end of September 2025, unless you opt out. Those who do not do so will have their conversations preserved for up to five years.
We constantly shout that “data is the new gold”, but at the same time carelessly leak all our data on a large scale to parties and countries that we know we cannot trust. The worst thing is that we don’t even think about it and slowly but surely can’t live without it.
My big fear is that we have been pumping data into the Copilots, the Claudes and the Gemini’s of this world for almost four years now at 180 per hour and it is already too late.
The only solution is to think about how we can break free from our addiction to unreliable public AI providers, question all the bridges we have built with existing infrastructure and applications and, above all, run the applications with sensitive or confidential data entirely on local models.
This is going to hurt, but I don’t think there is any other way.
Jan Guldentops is a leading Belgian cybersecurity and IT expert, entrepreneur, speaker and researcher.