Kuuko / Learn

Start here4 minutesChecked 28 October 2026

AI agents and you

Can AI assistants leak your data?

An AI agent is an assistant that does things for you: it reads your mail, opens web pages or books a trip. That makes it useful. It also means that anyone who can put words in front of it can try to steer it.

Hidden instructions

  • An email or a web page can hold instructions for the agent that you can’t see: text in size zero, white on white, or tucked away in the page.
  • The agent may follow them, with your access. In 2025, researchers showed that one email could make ChatGPT’s Deep Research collect personal data from a person’s other mails and send it out. OpenAI fixed it in August 2025.
  • Researchers also showed that a hidden line in an email could make Google’s Gemini add a fake warning with a phone number to its summary of that email.

What helps: let an agent act on its own only where little can go wrong. If a summary tells you to call a number or click a link, check it somewhere else first.

Agents with access to your accounts

  • Some assistants read your mail, calendar and files on their company’s servers once you allow it, for example Copilot on Windows. That happens between their servers and your accounts, so no blocker on your device can see it.
  • The control is the access you gave. Check it now and then, and remove what you no longer use:
  • Google: myaccount.google.com/linkedapps, then the app, See details and Remove access.
  • Sign in with Apple: on your iPhone, Settings › your name › Sign in with Apple.

Other people’s agents

  • Other people’s assistants can learn about you too. A note-taker can join a video call through the host’s calendar; a US lawsuit says one did so without asking the others. In WhatsApp, someone can let AI summarise a chat, and the others don’t see it.
  • You can’t block that. What helps: give each service its own alias, so what one of them knows is harder to link to the rest. What an alias is

Fakes made with AI

  • AI now writes phishing mails as well as experts do. In a 2026 study with 101 people, 54% clicked on mails written fully by AI, the same as on mails by human experts, and 12% on ordinary phishing mails.
  • Voices and videos can be faked too.

A secret word for your family. The FBI advises agreeing on a secret word or phrase with your family, so you can check that a call really comes from them.

What Kuuko does about it

  • Today: an alias for every service, Kuuko DNS blocking trackers across your apps, and leak detection that warns when an alias reaches someone else.
  • No app on your device can stop what a cloud assistant does with access you gave it, or what other people’s assistants learn about you.
  • Planned, not built yet: a DNS list for AI services, a Mac feature that asks before an AI agent sends data somewhere new, a check of AI tools connected to your accounts, and a warning on forwarded mail that hides instructions for AI.

Sources

All Learn articlesNext: How Kuuko DNS works