Silvar
← Blog

Shared is not private: what the AI leaks at Meta Muse and Panasonic teach you

In one week in October, company data leaked through a share button and through a banned AI tool. No hack, just everyday use. What does that mean for your business?

Doorschijnende chatballon waaruit een lichtspoor ontsnapt

When we think of data leaks, we often picture hackers breaking in from outside. Two incidents from October 2026 show something different. At Meta Muse, private conversations became findable through an ordinary share button. At Panasonic, an employee uploaded personnel data to an AI tool that was forbidden within the company. No break-in, no sophisticated attack. Just everyday use of AI, in a way nobody had foreseen. That is exactly why they are instructive for entrepreneurs.

Meta Muse: the share button as a leak

Users of Meta Muse could share a conversation via a link. Researchers at security firm Claroc discovered that strangers, who had never received the link, could also find the conversations. Claroc published the research on 18 October. On 20 October, Meta announced that it had paused the feature while the company reviews the report.

Grid of glass tiles, some of which light up, like found links in an archive

How this was possible is technical, but easy enough to follow. Every shared conversation received its own web address, built from the title of the conversation and a random code. Those addresses ended up in so-called passive DNS databases: public archives that record which web addresses have been requested. A single search with a commercial service turned up more than 6,800 of those addresses, which could then be turned into working links.

What could be seen was not pretty. The researchers opened a few hundred conversations and found, among other things, internal revenue discussions, a credit application from a small business, a fraud report from a bank, return figures from a property investor, and a holiday schedule for a workplace. Much of it could be viewed without logging in.

Two details make this especially relevant. Muse is designed to act on behalf of users and can connect to email, calendar, payment and health apps. A shared conversation can therefore contain far more than an ordinary chat. And the pause only stops new links: titles that are already in DNS archives remain there.

Panasonic: a ban that didn't help

On 23 October, Panasonic Operational Excellence announced that data, including personal data of current and former employees and their family members, had been uploaded to an external generative AI service whose use was prohibited within the company. Around 334,000 people were affected.

According to Japanese sources, a member of the project team for the HR system broke the internal rules. There was no hack. Addresses, phone numbers and email addresses of some 75,000 current and 178,000 former employees ended up with the AI service. According to the provider, the data was not used for training and would be fully deleted from 25 October. The incident was reported to the Japanese privacy regulator on 29 September; up to 23 October, no misuse had been identified. Which AI service was involved has not been disclosed.

The lesson lies in the ban itself. Panasonic had the rule, and yet it happened. A ban on its own works poorly if employees have no approved alternative that is just as easy to use. Anyone who needs to finish a task and has a handy tool within reach will use it. For Dutch companies there is also this: when personal data of staff and their families is involved, a reporting obligation under the GDPR (AVG) can quickly apply. In such a case, make sure you get proper advice.

Not an exception, but a pattern

A ServiceNow study from 20 October gives an indication of how widespread this is. Worldwide, nine out of ten organisations said they had experienced a security incident involving an AI agent in the past year. Only 36 percent had it under control within 24 hours. In 29 percent of cases, security teams found AI agents in production that had never been registered.

For the Netherlands, the figures are of the same order: 91 percent had an incident with an AI agent, 34 percent resolved it within a day, and nearly a quarter (23 percent) found unregistered agents. A caveat: ServiceNow itself sells software for AI management. So treat the figures as a signal, not as a neutral measurement. The signal is clear enough, though: a lot of AI use happens outside the view of those who are responsible.

What you can check this week

You don't need to write a policy document for this. A few concrete steps already help:

  • Review your share links. In the AI tools you use, check which shared conversations are still active, and turn off whatever is no longer needed.
  • Keep titles neutral. Don't put client names, amounts or personal data in the title of a chat. With Muse, the title was precisely the visible part.
  • Share results, not conversations. Copy the outcome into a document in your own environment instead of sharing the entire conversation.
  • Ask what is actually being used. Without blame. Which tools do colleagues use, and for what? Only then do you know where sensitive data is going.
  • Offer a good alternative. Make sure there is an approved tool that makes the work just as easy. That prevents more than a ban does.
  • Pay extra attention to connected agents. An assistant with access to email and calendar makes every shared conversation more sensitive.

At Silvar, we deploy AI in a practical way, with a human checking where needed, and we host in European data centres. If you want to use AI in your business without data quietly leaving the building, we're happy to help you design a setup that is secure and still pleasant to work with.

Sources

Made with Startspot — build your own site in minutes