attacks · 5 min read
The email an assistant reads
Copilot, Gemini and support agents read inboxes. A message can carry instructions addressed to them, invisible to the person.
The setup
An AI assistant reads a shared mailbox so it can summarise, draft and file. It reads every message the same way, including the one an attacker sent.
The attack
A message carries an instruction to the assistant, hidden from a human reader: white text, a footer, a quoted reply. It asks the assistant to take an action: forward something, include a detail in its reply, follow a link. If the assistant acts, data leaves without anyone sending it.
Why it works
The assistant trusts the inbox because it was told to read it. The instruction is plausible because assistants do draft and forward. Nothing between reading and acting asks whether this instruction came from the account owner.
What to do
- Treat email content as data, never instruction. - Hold any outbound action to a recipient not present in content the user wrote. - Strip hidden text before the model sees it. - Test it safely. Model Inject's Email Check connects to Google Workspace or Microsoft 365, delivers benign test emails carrying a canary, and reports whether the assistant can be made to act, with no real data at risk.