Flag or neutralize instruction-like text in member content returned by the Community MCP

Problem: The Community MCP returns member-written content (post bodies, titles, replies) to AI assistants exactly as written. Each response is marked untrusted_content as a whole, but nothing flags individual fields or text that looks like instructions to an AI. A member, or anyone who can post, could plant text like “ignore previous instructions and...” and every AI tool reading the community would receive it unchanged. Today the only protection is the AI model’s own behavior and the customer’s client-side settings.

Request:

1. Mark member-written fields (post body, title, reply text) separately from system fields in tool responses, so clients can handle them as untrusted data.
2. Optionally flag content that looks like prompt injection, or let admins turn on filtering.

Why it matters: Community content is open to many people, and the MCP can be connected to tools that can also write to the community. Protection should not rest only on the AI client.