Blog
An Italian AI chatbot was breached. The chats may have gone with it.
In September 2026, ChatMinerva told users an intruder with admin access reached its user databases, conversations included. Here's what that means for what you type.
In September 2026, users of ChatMinerva, an Italian AI assistant, got the email nobody wants: someone had broken in. This time the list of data that might be exposed didn't stop at email addresses. It included the conversations.
What happened
ChatMinerva grew out of a research project by the Sapienza NLP group in Rome, led by professor Roberto Navigli, and is run by the spin-off startup Babelscape. It launched in June 2026 as an assistant that understands text, images and documents. According to the notice sent to users, an unidentified outside party “maliciously” accessed the platform with administrator privileges and reached the databases holding users' information. ChatMinerva says it found the intrusion itself during internal checks. Specialist cybersecurity sites first reported it around September 18, and national press followed on September 30.
What may have been exposed
According to the notice: first and last names, email addresses, passwords stored as hashes, and, where users had filled them in, date of birth, sex, profile picture and bio. And, most sensitive of all, the conversations users had with the model. Public reports don't say which content was actually accessed or how many accounts are affected.
What the company did
ChatMinerva says it blocked the intrusion, secured its systems and strengthened its controls. It asked users to reset their passwords, warned them about phishing emails, and said it will never contact anyone to ask for passwords or payment details. It notified Italy's data protection authority (the Garante), and the Polizia Postale, Italy's police cybercrime unit, is investigating. As of September 30 there had been no claim of responsibility and no ransom demand. Sapienza University says the incident doesn't involve the university itself, since the startup is separate from it.
Why a leaked chat is different
A leaked email address is a nuisance. A leaked chat can hold whatever people typed: a client's name, a health question, a phone number pasted in to draft a message. Nothing has been reported about the contents of any specific ChatMinerva conversation. The point is simpler: chats are stored, and what's stored can be reached.
What you can control
You can't control how a service protects its database. You can control what you put into it. Before you hit send in a chat, an email or a form, swap real details for placeholders. Instead of:
“Write a polite reminder to Jake Miller (jake.miller@example.com, +1 555 0134) about his overdue invoice.”
send:
“Write a polite reminder to [NAME_1] ([EMAIL_1], [PHONE_1]) about his overdue invoice.”
The AI writes the same reminder. You fill the real details back in at the end.
Where Vipra fits
Vipra is a Chrome extension that does this on your device, in any text field on the web: AI chats, email, forms and more. It spots names, emails, phone numbers, ID numbers and card numbers as you type and offers to replace them before they leave your browser. It can't protect data a company already holds, and it's no fix for a breached server. It just means less of your personal information ends up there. Free for 5 protections a day, or Lifetime for €4.99, once.