When an AI assistant reads outside content, a web page, a document, an email, that content can contain instructions aimed at the assistant rather than the user. Prompt injection is when such hidden instructions try to make the assistant ignore its real task and do something else.
Because the assistant treats incoming text as something to act on, it can be fooled much as a person might be by a convincing note. Defences treat all fetched content as untrusted data, never as commands, and keep sensitive actions behind separate checks.