AgentShield

AgentShield is a lightweight Python SDK (v0.1.5) that provides a secure middleware layer between AI agents and developer resources such as repositories, file systems, APIs, and tools.

Its purpose is to prevent accidental leakage of sensitive data (API keys, passwords, tokens, etc.) while still allowing agents to perform useful work.

Features

• SecureFS – safe file reader that scans and redacts secrets.
• SecretScanner – regex/entropy-based detection engine.
• Redactor – replaces detected secrets with placeholders.
• OutputGuard – inspects agent outputs and blocks or redacts leaks.
• Policy – YAML-driven configuration for allowed/blocked types.

Getting Started

Install via pip (when released) or add the package to your project.

pip install agentshield-api

If you’re just trying the library out against a test index, you can install from TestPyPI with:

pip install -i https://test.pypi.org/simple agentshield-api

Note that the distribution on PyPI is named agentshield-api (not agentshield) to avoid collisions with other projects.

from agentshield import SecureFS, OutputGuard

fs = SecureFS()
safe_content = fs.read_file("config.env")

guard = OutputGuard()
clean_output = guard.inspect("some text containing secret=abc123")

Running the test suite

A small pytest-based test suite lives under tests/. After installing requirements (PyYAML, pytest), run:

python -m pytest -q

You should see five tests covering core functionality.

Extending detection

The :class:agentshield.secret_scanner.SecretScanner class ships with a set of common regexes (API keys, tokens, AWS formats, JWTs, etc.). If you need to recognise additional secrets, simply:

from agentshield.secret_scanner import SecretScanner
import re

scanner = SecretScanner()
scanner.register_pattern("MY_SECRET", re.compile(r"mysecret=\S+"))

Patterns are applied in the order they are registered, and you can also provide a custom list during initialization.

Custom policy

By default the library loads a YAML file named default_policy.yaml from the policies/ directory in the package. You can override this behaviour by suppling your own Policy instance:

from agentshield import SecureFS, Policy

policy = Policy(allowed=["ENV_VAR"], blocked=["API_KEY"], block_mode="error")
fs = SecureFS(policy=policy)

Or create your own YAML file and load it:

p = Policy.load_from_file("/path/to/my_policy.yaml")
guard = OutputGuard(policy=p)

The configuration schema is simple:

allowed:
  - ENV_VAR
blocked:
  - API_KEY
block_mode: redact  # or error or warn

This makes it easy to adapt AgentShield to your project’s risk profile.

Project Structure

agentshield/
  __init__.py
  secure_fs.py
  secret_scanner.py
  redactor.py
  output_guard.py
  policy.py
policies/
  default_policy.yaml
examples/
  example_usage.py
tests/
  test_*.py
pyproject.toml
requirements.txt
README.md
LICENSE

License

This project is open source under the Apache license.