Summary
This video from Fahd Mirza introduces a GitHub project called conversation-steganography, a tool that hides encrypted private messages inside AI-generated casual conversation text. Unlike traditional encryption — which makes it obvious that something is being protected — steganography conceals the very existence of a secret communication, and this tool applies that principle using a local language model as the carrier medium.
The technical mechanism involves two layers: AES-SIV encryption that cryptographically chains each message to prior ones (preventing tampering, deletion, or reordering), followed by feeding those encrypted bytes into a locally running GPT-2 model. Instead of sampling tokens normally, the model’s word choices are quietly steered by the encrypted data, producing output that reads as ordinary small talk while secretly encoding the message in the precise sequence of tokens chosen. The recipient’s copy of the same model reverses the process to recover the original text.
The video walks through the full setup on Ubuntu using Go (built from source) and Conda, downloading a 500 MB GPT-2 model that runs entirely on CPU without requiring a GPU. A live two-person simulation shows a secret phrase — “purple elephant dances under crimson moonlight” — transformed into innocent-looking cover text and successfully decoded on the other end. The project addresses a real privacy gap: even end-to-end encrypted messaging apps leave metadata traces revealing that a private conversation is occurring.
📺 Source: Fahd Mirza · Published August 01, 2026
🏷️ Format: Hands On Build
