Prompt injection
Text in content a model reads — a web page, an email, a document, a tool result — that the model follows as if it were an instruction from you.
What people get wrong
It cannot be fixed by telling the model to ignore instructions in data, because to the model there is no boundary between the two. The defences are architectural: least privilege, confirmation on actions that matter, and not mixing untrusted input with powerful tools.
A definition is not an understanding. This is taught properly in Why prompt injection cannot be prompted away, part of Prompt injection & guardrails.
Learn this properly