Conversation Steganography Hides Encrypted Messages Inside Innocent WhatsApp Chats
A Go-based tool wraps AES-SIV encryption with a local LLM that turns your secret messages into innocent-looking chat text.
Takeaways
- Open-source Go tool Conversation Steganography hides encrypted messages inside natural-sounding LLM-generated chat text
- Cover text can be pasted into WhatsApp, Signal, Telegram, or email without triggering content scanners
- Uses AES-SIV encryption plus PBKDF2 with 600,000 iterations and a cryptographic conversation chain
- Supports local Llama 3.2 1B/3B, Llama 3.1 8B via MLX on Apple Silicon, or GPT-2 via Transformers
- Both parties must share a secret phrase, conversation name, and identical model configuration
- Labeled a proof of concept, GPL-3.0 licensed, currently at over 1,200 stars
A new open-source project called Conversation Steganography is drawing attention on GitHub for a specific pitch: use a local language model to disguise encrypted messages as ordinary chit-chat you can paste into WhatsApp, Telegram, Signal, or email. The repo has crossed 1,200 stars and takes aim at the growing trend of client-side scanning by governments and platforms.
You type something like "meet me at the coffee shop at 3pm" and the tool generates something like "Hey, I was just thinking about that recipe you mentioned. It sounds amazing, especially the part about the fresh basil," which is what actually gets sent. Your friend's copy of the tool decodes it back to the original. The messaging platform, along with anything scanning it, only ever sees the innocuous cover text.

Encryption first, language model as camouflage
Under the hood, the pipeline chains familiar cryptographic primitives with an LLM bolted on as the carrier. Messages are encrypted with AES-SIV, which provides both confidentiality and integrity. Every message is chained so each carrier authenticates the sender, conversation ID, message index, and the hash of all previous messages. The shared phrase is stretched with PBKDF2-HMAC-SHA-256 using 600,000 iterations. The AI model runs 100% locally with no API calls, no cloud, and no telemetry. Conversation state is encrypted at rest with AES-GCM.