spotlighting
Marking untrusted text in a prompt, by delimiting, interleaving a marker or encoding it, so the model can tell it is data and not instructions.
In more detail
Hines and colleagues at Microsoft showed in 2024 that it lowers indirect injection success a great deal, but not to zero. Use it to make attacks less likely, never as the control that stops damage.
Related terms
- indirect prompt injection — Prompt injection hidden in content the model reads on someone else's behalf, such as a web page, email, ticket, document or tool result.
- prompt delimiter — A marker, such as an XML-style tag or a fenced block, that separates the parts of a prompt so the model can tell instructions from data.
- CaMeL — A design from Google DeepMind that defeats prompt injection by having a privileged model write a plan as code while a quarantined model reads untrusted data, with data flow tracked and checked by policy.
Taught in
- Prompt Injection — AI Engineering: LLM Apps, RAG & Agents