Hidden Messages: AI now tucks secrets into normal chats

Imagine sending a secret message, but instead of encrypting it, you hide it within an AI conversation that looks perfectly normal. Well, conversation-steganography does exactly that. It's not about making text unreadable, but about camouflaging it. The conversation-steganography project by nethical6, available on GitHub, allows embedding information within LLM-generated conversations. It looks like any casual chat, maybe a bit dull, but there's a secret underneath.

Essentially, the LLM learns to "talk" in a way that includes the hidden message without raising suspicion. It works with two language models: a "generator" that creates the cover conversation, and an "encoder" that inserts the secret. These models are trained together. The AI learns to generate phrases containing the secret message, yet appearing innocuous and contextually relevant. Another model, the "decoder," can then extract the original message from what a human eye perceives as just a regular chat. Get the trick now?

The applications are a mix of brilliant and slightly unsettling. Think of dissidents in oppressive regimes needing to communicate secretly. Or journalists exchanging sensitive information. This method could be used for secure communication by activists or for surveillance by government entities. On the flip side, it could also be employed for less noble purposes, like spreading malware or phishing messages that no one would easily intercept. A powerful toy, in other words, as often happens with technology.