// GLOSSARY

Indirect prompt injection

An instruction planted in third-party content the model reads (a document, web page, email or knowledge-base record) being processed as though it were the user's request. The attacker never talks to the model; they write to a source the model trusts.