Multi-Agent Firewall Architecture for Privacy Protection of Sensitive Data in Interactions with Language Models

2026-03-30 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, Software Development & Engineering · Depth: Expert, extended

Summary

This paper introduces an open-source, multi-agent firewall architecture designed to protect sensitive data during interactions with Large Language Models (LLMs). The system addresses data leakage risks in both web-based and programmatic LLM usage by combining a browser extension and a transparent Man-in-the-Middle (MiTM) proxy for HTTP(S) and WebSocket traffic interception. Its core features include multimodal analysis, customizable detection via a hybrid pipeline of deterministic and LLM-driven semantic analysis, proprietary code leakage prevention, and granular risk enforcement. The architecture supports flexible deployment models, from standalone workstations to centralized local servers, and employs a sanitization-first workflow to prevent sensitive data exposure to detection models. Evaluation results demonstrate F1 scores up to 94.93% on optimal configurations, balancing detection depth, computational cost, and latency.

Key takeaway

For AI Architects and MLOps Engineers concerned with data privacy in LLM deployments, this multi-agent firewall offers a critical solution. You should consider implementing this local-first architecture to secure both web-based and programmatic LLM interactions, eliminating the privacy paradox of cloud-based scanning. Prioritize configurations that balance detection accuracy (e.g., fine-tuned models achieving 94.93% F1) with acceptable latency for your operational needs, ensuring sensitive data remains within your controlled environment.

Key insights

A local-first, multi-agent firewall architecture provides robust data leakage prevention for LLM interactions without modifying applications or externalizing trust.

Principles

• Hybrid detection balances speed and depth.
• Sanitization-first workflow preserves privacy.
• Modular design enables extensibility.

Method

The system intercepts LLM traffic via a browser extension and MiTM proxy, then routes it through a DAG-orchestrated multi-agent pipeline for hybrid detection (deterministic, NER, LLM-based semantic analysis) and policy enforcement, followed by sanitization.

In practice

• Deploy a local proxy for API traffic.
• Use browser extensions for web LLM interfaces.
• Fine-tune small models for PII detection.

Topics

• Large Language Models
• Data Leakage Prevention
• Multi-Agent Systems
• Privacy-Preserving AI
• Prompt Security
• Network Proxy
• Named Entity Recognition

Code references

• unslothai/unsloth
• rapidfuzz/RapidFuzz
• BerriAI/litellm
• langchain-ai/langgraph
• microsoft/presidio

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Security Engineer, MLOps Engineer, AI Architect

Related on AIssential

• Multi-Agent Firewall Architecture for Privacy Protection of Sensitive Data in Interactions with Language Models — A multi-agent firewall architecture secures LLM interactions by intercepting traffic and preventing data leakage through hybrid detection.
• Agents That Know Too Much: A Data-Centric Survey of Privacy in LLM Agents — LLM agents' extensive data interaction creates complex privacy risks, requiring a data-centric governance approach.
• Why Local-First AI Agents Are the Future (And Why It Matters for Your Privacy) — AI agents, while increasingly capable of deep system access, raise critical privacy concerns as most cloud-based architectures upload sensitive data like full screen screenshots, passwords, and financ…
• Trustworthy Agent Network: Trust in Agent Networks Must Be Baked In, Not Bolted On — Trust in LLM-based agent networks must be architected into the system's core design, not retrofitted.
• Five labs, five minds: building a multi-model finance drama on small models — Building multi-agent simulations with heterogeneous small models requires robust serving layers and careful information management.
• Stop Choosing Between Local and Cloud LLMs: A Field Guide to Hybrid Patterns — Hybrid local-cloud LLM patterns enable combining local privacy with cloud reasoning through a three-axis design framework and five common architectures.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.