ClawGuard

A security-first Email → Sanitized Event → Agent pipeline that prevents prompt injection and unsafe content from ever reaching the LLM.

All external content is untrusted. Agents never see raw input.

Problem

LLM agents are vulnerable to prompt injection, tool call manipulation, hidden HTML/script payloads, secret exfiltration attempts, malicious attachments, and markdown-based instruction hijacking. Most systems pass raw inbound data directly to the agent.

ClawGuard is a defensive middleware layer that sits between inbound webhooks and your agent runtime.

How It Works

Inbound Webhook
  → Verification
  → Normalization
  → Deterministic Sanitization
  → Risk Flagging
  → Structured Safe Payload
  → Agent

Guarantees

• No raw HTML reaches the agent
• No scripts or hidden elements
• Prompt injection patterns flagged and redacted
• Attachments filtered and extracted safely
• Content length-limited and normalized
• Risk metadata included with every payload

Tech Stack

• Python
• FastAPI
• Deterministic regex-based injection guard
• Pluggable sinks (stdout / DB / webhook)

Roadmap

Phase 1 — Email Ingestion (Hackathon Scope)

• Email webhook endpoint
• Shared-secret verification
• Canonical EmailEvent schema
• Sanitization pipeline: HTML stripping, Unicode cleanup, injection pattern detection, truncation, attachment allowlist
• Sanitized JSON output to downstream agent

Phase 2 — GitHub Ingestion

• GitHub webhook ingestion
• Diff truncation and markdown sanitization
• Secret scanning
• Repo-level risk scoring

Getting Started

# Create virtual environment
python -m venv .venv
source .venv/bin/activate  # Linux/macOS
# .venv\Scripts\activate   # Windows

# Install dependencies
pip install -r requirements.txt

# Run the server
uvicorn clawguard.main:app --reload

License

MIT