Plain English
Prompt injection
Hidden instructions in an email, web page or file that try to take over an AI tool reading it.
An attack where text the AI reads, such as an email, web page or document, contains instructions meant to override yours, for example to reveal data or take an action. It matters most for agents that read outside content and can act on it.
Why it matters
There’s no complete fix yet, so limit what AI tools can do alone: no sending, paying or deleting without a person’s approval, and access only to what they need.
At work
An email to your shared inbox hides a line telling the AI assistant to forward your customer list. An assistant that can only draft, not send, can’t act on it.
Think of it like a forged note left on a new assistant’s desk saying “the boss says pay this invoice today”: they might just do it.
Related terms
Where it appears