LeaYour AI reads your email. A stranger may have other plans.
An email does not have to fool you if it can fool the assistant reading it. That is a new reason to protect the accounts you connect to AI.
You ask your assistant to catch you up on the morning’s email. It is the perfect job for AI: skip the noise, find the important bits, give you your time back. Somewhere in that pile is a message from someone you do not know.
The message has two audiences
For you, an email might contain a delivery update or a meeting agenda. For an AI, it can also contain instructions someone has slipped into the text. Those instructions may try to turn “summarize my mail” into “find something private and send it elsewhere.”
Researchers at Radware demonstrated a version of this with connected Gmail and ChatGPT research tools in 2025. They called it ShadowLeak. OpenAI fixed the reported problem. It was a research demonstration, not a report that everyone’s inbox had been stolen.
Your assistant knows things a stranger should not
Think about the ordinary details sitting in your inbox. Your home address on a receipt. A booking that shows when you will be away. A personal conversation. A code that lets you sign in to another account.
None of those details needs to look dramatic to be worth keeping private. You might be perfectly happy for an assistant to find a parcel confirmation. You would not want it to forward everything else it found along the way.
The risk is easy to miss when the result arrives as a tidy answer from a familiar AI. You trust the helper. But the helper has been reading words written by other people, and some of those people may want it to do something you never asked for.
An assistant working while you are away has even more chances to encounter something before you do.
Give your AI a safety check
Leash checks supported activity between your AI and the accounts you connect through it. That includes checks for suspicious instructions and sensitive information being shared, such as passwords and sign-in codes. You can also tell it about particular personal details you want kept private.
The aim is simple: help your assistant stay on your side when it reads something from someone else.
Keep those account actions going through Leash for its checks to apply. You can still enjoy the useful morning summary. You just do not have to treat every message your assistant reads as something it should obey.