Willow External Guardv1.0.0

A customer support chatbot ingests messages from external channels like Telegram and Discord. The guard skill scans each inbound message for prompt injection and role hijack attempts before processing by the LLM, blocking malicious inputs that could hijack the bot's behavior.

A research assistant tool fetches web pages for summarization. The skill wraps raw HTML content in sandwich defense markers and scans for indirect injection, preventing embedded instructions from altering the summary or leaking system prompts.

During a Windows to Linux migration, unknown provenance files from the corpus are scanned before LLM-based classification or summarization. Suspicious patterns trigger user confirmation to avoid processing malicious content.

A multi-agent system spawns sub-agents for task decomposition; their outputs are scanned for role hijack or approval bypass before being acted upon. High-risk patterns are blocked entirely to prevent agent compromise.

Offer the guard skill as an optional security layer for existing LLM-based products. Integrate via API for per-call scanning and wrapping, charging per scan or per thousand tokens.

License the skill to enterprises deploying custom LLM pipelines for internal knowledge bases, customer-facing chatbots, or document processing. Provide premium support and custom rule sets.

Maintain the core guard script as open source, but offer a cloud monitoring service that aggregates and analyzes guard events across deployments, providing threat intelligence dashboards.