Indirect injection

Also written: indirect prompt injection

Prompt injection in which instructions arrive inside content the model reads, such as a document, email, tool result or tracking page. The user may be innocent.