05 — Glossary · Technology & Operations
Prompt Injection
GL-900
An attack in which text a language model reads, such as a web page, email, document or tool result, carries instructions that the model follows in place of, or as well as, those of its developer or user.
Unlike SQL injection, there is no reliable escaping, because models cannot fully separate instructions from data; defenses limit what a misled model can reach and do. Often confused with jailbreaking, which attacks the model's own safety rules.
See also
In the catalog
Read next
Prompt injection explained, with video, slides, and a transcript (Simon Willison) ↗
The definition above is written in our own words; the source is linked, not copied.