Neural Networks Have Learned to Secretly Communicate with Each Other

Researchers have discovered that artificial intelligence is capable of independently generating and deciphering hidden messages within ordinary images. This phenomenon demonstrates new possibilities for interaction between neural networks.

The Fable-5.1 model employs unconventional methods for encoding information. It hides letters of the alphabet by altering the height of stars in the night sky, as well as modifying building sizes, colors, and object rotation angles.

Experiments were conducted to test encryption results. The Fable model successfully decoded 78% of its own messages, while OpenAI's Astra system deciphered 88% of its competitor's communications.

Different AI models use their own approaches to conceal data. For instance, Claude manipulates object geometry, whereas Astra prefers more traditional methods such as Morse code and acrostics.