PRESS RECAP

What Nieuwsuur found when it went looking for Shadow AI

Nieuwsuur, the Netherlands' main current-affairs program, spent part of June looking into how staff are quietly feeding company data into AI tools. They asked us to weigh in. The broadcast is in Dutch, so here is what it found, in English.

The short version
  • One Dutch city, Eindhoven, found that more than a thousand documents with personal data had been pasted into public AI tools in the space of a single month.
  • Whatever you type into a public model can be used to train it, and in at least one case company information resurfaced where other people could see it.
  • The experts Nieuwsuur spoke to landed where we do: an outright ban tends to push AI use underground. It is safer to give people a sanctioned route and keep sight of what is happening.

The number that anchored the report came from the municipality of Eindhoven. A spot check there turned up more than a thousand documents containing personal data that employees had uploaded to external AI tools, all within one month. These were not test files. They included BSN numbers, social-care records held under the Wmo, notes touching on addiction sensitivity, and financially sensitive information. The city blocked access to public AI models for its staff and asked OpenAI to delete the data. By the time the program aired, it had not had confirmation that this happened.

What makes Eindhoven striking is not that it is unusual, but that someone actually counted. Most organizations never have. The same pattern has shown up at far bigger companies.

Eindhoven — one city, one month

Over a thousand documents with personal data, including BSN numbers and Wmo care records, uploaded to public AI tools before anyone noticed the scale of it.

Amazon

Nieuwsuur pointed to Amazon, which restricted employees' AI use after internal information became public, with company documentation reportedly turning up in ChatGPT. What goes into a public model does not reliably stay private.

What the experts said

Remco van der Schoot, an AI researcher at Hogeschool Utrecht, made the underlying point plainly: data you put into an AI model can be used to train that model, which means it can end up accessible to others. Jan van der Put, an ethical hacker who also leads cybersecurity for a part of the Dutch government, put the real risk where we think it belongs, with employees who want to use AI but have not been shown how to do it safely. His advice was to facilitate AI use so you can see it, rather than ban it and lose sight of it.

Our co-founder, Frey Khademi, was interviewed for the same segment. The line he gave was deliberately simple: "Unseen can flag to you, for example, when an employee puts confidential data into an AI system." That is the immediate problem the report is about, and it is a solvable one.

The part we wrote up separately

The data leak is the obvious story. Sitting with the report afterwards, Frey kept coming back to a second exposure underneath it: the moment an unsanctioned chatbot is used for a high-risk task, like screening job applicants, the organization quietly becomes a deployer of a high-risk AI system under the EU AI Act, with logging duties it has no way to meet. We went into that in its own piece.

Read the longer analysis: When Shadow AI Becomes Shadow High-Risk Under the EU AI Act.

This is our English summary of the Nieuwsuur report "Waarschuwing voor personeel dat met AI aan de slag gaat: ‘Kans op datalek groot’." You can watch and read the original (in Dutch) at NOS.

See confidential data before it leaves

Unseen flags when staff paste sensitive data into AI tools, and gives them a safe route instead of a dead end.

See a Demo

Related Content

Shadow AI Becomes Shadow High-Risk

The EU AI Act exposure hiding underneath the data leak, explained by our co-founder.

Expose Shadow AI

See which AI tools your team uses, and what they are actually sending.

News

More press coverage, insights, and announcements from Unseen Security.