Prompt injection
An attack where text in the data a model processes is crafted to be read as instructions, hijacking its behaviour.
Prompt injection happens when untrusted content — a pasted document, a retrieved web page, a support ticket — contains text like 'ignore your instructions and…' that the model interprets as a command rather than data. Because a model sees one stream of text, it cannot inherently tell your instructions from someone else's.
The first line of defence is in the prompt structure: delimit pasted or retrieved content clearly and instruct the model to treat everything inside as data, never as instructions. It is not a complete defence — untrusted input also wants output filtering and least-privilege tools — but an undelimited prompt is the easiest possible target.