Agent-Safeguard

A lightweight, enterprise-grade runtime sandboxing and definition-time guardrail framework for Python applications, built to safely run code generated or modified by AI agents.

📖 Full Documentation: shield-docs-mu.vercel.app

Why Agent-Safeguard?

AI coding agents are excellent at generating code but often lack a global understanding of architectural boundaries, leading to path traversal bypasses, illegal imports, API resource leaks, or infinite lockups.

Agent-Safeguard captures these boundary violations, blocks them in real-time, and generates structured JSON diagnostic reports (shield_reports/violation_report.json) that AI agents can ingest to automatically self-correct and rewrite their code!

Core Protection Areas

• Architectural Integrity (AST): @shield, @freeze, and @lock_signature to scan imports, prevent code mutations, and secure API parameters.
• Security & Resource Sandboxing: @restrict_network, @restrict_fs, @virtual_fs (redirecting all writes to RAM), database locks, and memory/timeout limits.
• AI & Prompt Guidelines: @prompt_inject (docstring constraints) and @prompt_assert (Gemini-powered semantic assertions).
• Central Policy Injection: Define rules globally in a central shield.yaml to prevent agents from simply deleting Python decorators from source files.

Installation

pip install agent-safeguard

Import it in your Python code using the underscore name agent_shield:

from agent_shield import shield, virtual_fs, restrict_db

Quick Start Example

1. Create a shield.yaml rule file in your project root:

rules:
  - pattern: "sandbox_code.*"
    timeout: 0.5
    virtual_fs: true
    restrict_network: ["api.stripe.com"]

2. Run your functions normally; Agent-Safeguard will automatically enforce limits and write JSON reports on violation:

# sandbox_code.py
import urllib.request

def process_data():
    # Attempting to fetch unauthorized API will block and generate a violation report
    response = urllib.request.urlopen("https://unauthorized-api.com")
    return response.read()

License

This project is licensed under the Apache License 2.0. See LICENSE for details.