Firebreak

Why Firebreak?

Python has no built-in sandboxing. Running untrusted code — whether from users, plugins, or AI — means giving it full access to your system. Firebreak fixes this by executing decorated functions inside lightweight Firecracker microVMs.

How It Works

from firebreak import firebreak

@firebreak(net=False, fs="ro:/data")
def process_untrusted_input(data: str) -> dict:
    # This runs in an isolated microVM
    # No network access, read-only filesystem
    return parse_and_transform(data)

# Call it like any normal function
result = process_untrusted_input(user_data)

At import time, the decorator replaces the function with an RPC stub. When called, arguments are serialized and sent to a warm microVM from a pool. The result (or exception) is returned transparently.

Key Concepts

Performance

This system trades raw call speed for strong isolation. That's the point.

Rule of thumb: if the function's own work takes <1ms, sandboxing will dominate. If it takes 10–100+ ms, overhead is usually acceptable.

Use Cases

• AI-generated code — Run LLM outputs safely
• User plugins — Let users extend your app without risk
• Data processing — Parse untrusted files in isolation
• Policy enforcement — Sandbox security-sensitive operations
• Multi-tenant execution — Isolate customer workloads