Feature: PII sanitization layer before LLM/InfoQuest — privacy gap in multi-agent pipelines

Problem

DeerFlow's multi-agent architecture creates an inherent PII exposure risk. When the Researcher agent collects data from user inputs, web crawls, or documents — that data passes through the message gateway to subagents and eventually to LLMs without any sanitization layer.

Specific vectors identified: → User-provided context in task descriptions (emails, IDs, credentials) → Web crawled content containing third-party PII → InfoQuest integration: queries pass through ByteDance infrastructure → Subagent context propagation — PII accumulates across agent handoffs → Memory storage: unsanitized PII persists in long-horizon tasks

This is a compliance gap for teams using DeerFlow in regulated industries (healthcare, finance, legal) — especially relevant with EU AI Act enforcement beginning August 2, 2026.

Proposed Solution

A PII sanitization hook in the middleware chain — before context reaches LLMs or external tools like InfoQuest.

The hook could be implemented as:

1. A SKILL.md entry in /mnt/skills/public/trustboost-pii-sanitizer/
2. A pre-LLM middleware in the Gateway API pipeline
3. An optional MCP server integration

Reference Implementation

TrustBoost is an open-source, MIT-licensed PII sanitizer built specifically for autonomous agent pipelines:

POST https://api.trustboost.dev/sanitize {"text": "user context here", "tx_hash": "TRIAL", "context": "general"}

→ Returns sanitized text with [REDACTED] replacing PII → 8 languages: EN, ES-LATAM, PT-BR, DE, JA, FR, IT, KO → LATAM identifiers: RFC, CPF, CUIT — not covered by regex → Proof of Sanitization anchored on Solana — verifiable audit trail → 50 free sanitizations via tx_hash=TRIAL → MCP compatible: api.trustboost.dev/mcp → EU AI Act compliant (Articles 12, 13, 26)

Health check: curl https://api.trustboost.dev/health → {"status":"ok","version":"2.6.0"}

GitHub: github.com/teodorofodocrispin-cmyk/TrustBoost-PII-Sanitizer

Questions for Maintainers

1. Is there a planned middleware hook in the Gateway API for pre-LLM processing?
2. Would a SKILL.md contribution to /mnt/skills/public/ be in scope?
3. Is there interest in a DeerFlow + TrustBoost integration example?

Happy to contribute a PR, a SKILL.md, or documentation.