austin-weeks/miasma

What it solves

It prevents AI companies from scraping public websites for training data by trapping web crawlers in a loop of "poisoned" training data. This protects website owners from having their work stolen for model training without consent.

How it works

Users deploy a Miasma server and use a reverse proxy (like Nginx) to route specific traffic to it. By embedding hidden links on their website that are invisible to humans but visible to scrapers, the owner can lure bots into a "trap." Once inside, Miasma serves the bot an endless stream of poisoned data from a source (such as the poison fountain) and provides self-referential links that keep the bot trapped in a cycle of consuming low-quality or misleading data.

Who it’s for

Website owners and content creators who want to discourage AI scrapers and protect their intellectual property from being used in LLM training sets.

Highlights

• Low Resource Usage: Built in Rust for high speed and a minimal memory footprint.
• Bot Trapping: Uses hidden links and self-referential loops to keep scrapers occupied.
• Poisoned Data: Proxies training data designed to degrade model quality.
• Traffic Metrics: Includes a built-in SQLite database to track request counts per unique User-Agent.
• Flexible Deployment: Can be run as a binary, a library, or via Docker.