Why should model output be treated as untrusted data?

Validation follows the entire data flow because a value safe for plain text may become dangerous when later embedded in HTML, Markdown, a query, a log, or another prompt. Generated artifacts carry provenance and trust level so downstream services do not accidentally promote them to authoritative data. Failures are explicit and observable rather than repaired into a different action without user awareness. Security evaluation exercises second-order paths where one model's output becomes another model's instruction or a privileged tool input.