Cross-platform epistemic access control for AI coding assistants - Claude Code, Cursor, GitHub Copilot

Epistemic Guardrails for AI Agents

Controlling what knowledge systems can access, retain, and operate on.

A contribution from Theios Research Institute, Inc.

Disclaimer

This software is provided as-is with no warranty. The authors are not responsible for data loss, security breaches, or any damages resulting from the use of this software. This tool is an aid, not a guarantee of protection. Users are responsible for verifying their own security configurations and compliance requirements. This is not legal advice.

The Problem

AI coding assistants like Claude Code, Cursor, and GitHub Copilot can retain conversation history through memory features. While useful for continuity, this creates a critical challenge:

How do you work on sensitive projects (proprietary research, pre-patent work, trade secrets) while using memory for routine development?

The answer: Epistemic Guardrails - a framework that enforces information boundaries based on project sensitivity.

Four Layers of Protection

1. Session-Start Guard - Warns the AI about sensitive directories when memory is enabled (SessionStart hook)
2. PreToolUse Hook - Blocks file access to sensitive directories during active sessions (hard enforcement)
3. Path + Keyword Detection - Identifies sensitive projects by directory path and naming patterns
4. Outbound Action Guard - Blocks pushes, publishes, and deploys to unauthorized destinations (Bash tool interception)

Overview

Epistemic Guardrails provides a unified framework for controlling info