Harvest

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.75). Yes—Step 2 fetches public URL/YouTube transcripts via runtime fetchers (WebFetch / fabric -y) and Step 7 always ingests the original input as a KNOWLEDGE note via _HARVEST (via harvest.ts), which would include outsider-authored free-text from the fetched content into the agent’s LLM context through the Knowledge Archive note.