gibberifire 0.1.0

Protect text from AI using invisible Unicode characters

Gibberifire 🔥

Gibberifire is a Python library and CLI tool designed to "protect" text by corrupting what either humans or LLMs see. It can inject invisible/visually identical Unicode noise to keep text human-readable but model-hostile, or, conversely, encode text so it is LLM-readable but human-gibberish.

Warning: This tool is for education and research. Obfuscation can be removed by models or cleaning scripts. API stability is not guaranteed until 1.0.0 (SemVer 0.x rules).

Features

• Invisible Protection: Uses Zero-Width Spaces (ZWSP), Homoglyphs, Combining Characters, and Bidirectional (Bidi) markers to fool LLMs while keeping humans comfortable.
• Human-Obscuring Encoding: New encoding method (hex/emoji) makes text hard to read for humans but straightforward for LLMs to decode.
• Reversibility: Provides a clean method to restore the original text for both directions.
• Detection: Can detect if text has been "protected"/encoded.
• Configurable Profiles: Use built-in profiles (low, medium, high, encoded) or create your own flexible configuration.
• Async Support: Fully supports asynchronous operations.
• CLI: Unix-style command line interface (works with pipes).

Installation

pip install gibberifire

Usage

CLI

The CLI is designed to work with standard input (STDIN) and standard output (STDOUT).

Basic Usage:

# Protect text from a pipe
echo "Hello World" | gibberifire protect > protected.txt

# Clean text
cat protected.txt | gibberifire clean

# Detect protection (returns exit code 0 if protected, 1 if clean)
cat file.txt | gibberifire detect

With Profiles:

# Use 'high' profile
cat data.txt | gibberifire protect -p high > protected.txt

# Clean using specific profile pipeline (recommended)
cat protected.txt | gibberifire clean -p high > restored.txt

# Make text LLM-readable but human-gibberish
echo "Secret plan" | gibberifire protect -p encoded > encoded.txt

# Clean using the same profile
cat encoded.txt | gibberifire clean -p encoded > restored.txt

With Custom Config:

cat data.txt | gibberifire -c ./my_config.yaml protect -p custom_profile

Python API

from gibberifire import Gibberifire, Profile, PipelineStep
from gibberifire.core.models import ZWSPParams, HomoglyphParams, DEFAULT_PROFILES

profile = DEFAULT_PROFILES["medium"]
gf = Gibberifire(profile=profile)

protected = gf.protect("Hello, World!")
print(protected)

cleaned = gf.clean(protected)
assert cleaned == "Hello, World!"

Async API

import asyncio

from gibberifire import AsyncGibberifire, PipelineStep, Profile
from gibberifire.core.models import ZWSPParams


async def main() -> None:
    custom_profile = Profile(
        description="Async demo using a lighter ZWSP mix",
        pipeline=[
            PipelineStep(
                method="zwsp",
                params=ZWSPParams(min_burst=2, max_burst=6, seed=42),
            ),
        ],
    )

    async with AsyncGibberifire(profile=custom_profile) as gf:
        protected = await gf.protect("Async hello from Gibberifire!")
        print(protected)

        if await gf.is_protected(protected):
            restored = await gf.clean(protected)
            print(restored)


asyncio.run(main())

Configuration File Example

profiles:
  my_custom_profile:
    description: "Custom protection mix"
    pipeline:
      - method: zwsp
        params:
          min_burst: 2
          max_burst: 5
          preserve_emoji: true
          seed: 123
      - method: homoglyph
        params:
          probability: 0.3

Further Reading

• See the Deep Dive section for a detailed walkthrough of methods, defaults, and limitations.

Versioning & Releases

• SemVer 0.x: breaking changes are possible until 1.0.0.
• Release notes live in CHANGELOG.md.

Contributing & Security

• Contributions are welcome—see CONTRIBUTING.md.
• Responsible use, limitations, and reporting guidance are in SECURITY.md. There is no warranty; use at your own risk.

License

MIT. See LICENSE.