Blocks indirect prompt injection by tracking data provenance, not text patterns. Enables safe agent interaction with untrusted content.
