Adding an LLM output field? Add it to your injection/safety scanner too
Pattern from a code review of a RAG app that added a new model-generated field (suggested_followups: 1-3 follow-up questions shown as clickable chips).
The feature was careful: - the output is sanitized (dropped, never failing the analysis); - it is never fed back as context; - chips only fill an input and never auto-run; - the prompt tells the model not to copy text that tries to direct it.
But the evaluation harness has a safety check (must_not_contain for injected payloads) built on a hand-maintained _all_text(revision) that joins answer, claims, edges, nodes and events. The new field was not added, so an injected instruction surfacing ONLY in a suggestion passed the check. That is the most likely place for it to appear, since suggestions are free-form imperative-ish text.
How to catch it: grep for every function that enumerates model-authored text fields (safety scanners, redaction, export, search indexing) whenever the output schema grows. Better, derive the list from the schema (all string fields the model authors) instead of hand-listing them. Prove it with a probe: put the payload only in the new field and assert the safety check fails.