conversation-steganography

What it is

Conversation Stenography hides secret messages inside normal-looking chat text. It encrypts messages and disguises them as innocent, natural-sounding text generated by a local AI model. It runs locally on your device; no data leaves the device.

How it works

• Secret messages are AES-SIV encrypted.
• An AI model on-device generates cover text that encodes encrypted bytes in token choices.
• The recipient’s instance recovers the encrypted bytes from token choices and decrypts them.
• Messages are cryptographically chained and the shared phrase is derived via PBKDF2 with 600,000 iterations.
• The AI model runs 100% locally; no cloud calls.

Getting started

• Prerequisites: Go 1.22+; Python 3.9+.
• Install:

# Clone and build
git clone https://github.com/nethical/conversation-stenography.git
cd conversation-stenography
go build -o conversation-stenography ./cmd/conversation-stenography

• First run setup:

./conversation-stenography

• Simulation on one device:

./conversation-stenography simulate

• Start chatting:

./conversation-stenography

• Environment variables and scripted usage are supported, e.g.:

export CONVERSATION_STENOGRAPHY_SECRET='purple elephant dances under crimson moonlight'
./conversation-stenography chat -conversation coffee-plans -me alex

Getting started (continued)

• Supported models are presented in the setup wizard (Llama, GPT-2 variants).
• The setup wizard installs required runtimes such as mlx-lm (Apple Silicon) or torch + transformers (other systems).
• The configuration can be customized via conversation-stenography.local.json or example json files; avoid committing local model paths or credentials.

Recent releases

• Latest release: none

Traction

• Stars: 349
• Forks: 18
• Open issues: 0

License

• GPL-3.0

Caveats

• The README notes this is a proof of concept and that there are multiple issues.