Researchers Propose Multi-Agent Firewall to Block Sensitive Data Leaks in LLM Interactions

A preprint posted to arXiv describes a browser extension and proxy system that intercepts all HTTP, HTTPS, and WebSocket traffic to and from large language models. The architecture routes requests through a multi-agent pipeline that combines deterministic pattern detectors with LLM-based semantic analysis to catch proprietary code, personal data, and other sensitive content before it leaves the user's environment. In evaluation, the best configuration reached an F1 score of 94.93 percent on data-leakage detection. The authors position the system as an open-source privacy layer for both web and programmatic LLM use.