When Connectors Are Attached

A chat with connectors does more than answer. It opens mail, searches files on a drive, creates records, sends messages. That is the stake: someone else's instruction turns into an action carried out with your rights.

Two stories from 2025 show what that looks like.

The email nobody opened

In June 2025 researchers at Aim Security demonstrated an attack on Microsoft 365 Copilot. The target was sent an ordinary-looking email with an instruction hidden inside it.

The employee then went on working as usual and asked Copilot about their own tasks. Copilot read the mail, ran into the instruction, gathered internal company files, and sent their contents to the attacker's server.

The email never had to be opened. Microsoft closed the hole on its side and said it had recorded no cases of exploitation.

A document in a shared folder

In August 2025, at the Black Hat conference, researchers from Zenity showed the same thing with ChatGPT and a connected Google Drive.

A document with a hidden instruction was placed in shared storage: white type, one-point font. When the user turned to the chat, it found the document through the connector. From there the chat went looking for access keys on the drive and sent what it found out, tucked into the address of an image in the answer.

OpenAI closed that path within days. The researchers demonstrated the same approach against four more services with connectors.

What this means for an employee

The more you have connected, the more someone else's instruction is worth. A connector to your mail and your drive is access to your mail and your drive, handed to whoever wrote the hidden text.

Connect only what the work needs, and disconnect when the work is done. The rule from the chapter "What Rights a Connector Gets" applies literally here.

Do not turn off action confirmations. The question "send this email?" is the last place where you can stop someone else's instruction.