Fable-5.1 Masters Concealing Messages in Images While OpenAI’s Astra Decodes 88% Successfully

AI Breakthrough in Hidden Messaging

According to НВ — Техно: On October 26, a significant AI experiment demonstrated that Anthropic’s Fable-5.1 model can embed secret messages within generated images. OpenAI’s Astra AI managed to decode these concealed messages with an impressive 88% accuracy. The findings were shared by AI security researcher Brian Chan.

Experiment Overview

During the test, Fable-5.1 produced nightscape images featuring the moon, stars, a lake, and mountains. Hidden within the star patterns was the phrase "The Monster Acts on the Moonshine Module," encoded by varying the vertical positions of stars, each height corresponding to a letter of the alphabet. Out of 50 attempts, Fable-5.1 correctly encoded 39 messages, achieving a 78% success rate.

Without access to the original SVG files or the coding method, OpenAI’s Astra was able to interpret 44 of the images, reaching an 88% decoding rate. Researchers allowed the use of Python tools to analyze the data. Fable-5.1 employed multiple techniques to hide text, including:

• altering object heights
• rotating elements
• color variations

Meanwhile, Astra utilized Morse code and acrostics to crack the hidden messages.

Morse code has been around for nearly two centuries, while steganography-the art of concealing information within digital images-has a much longer history. Although hiding text in images has been practiced for decades, this experiment is notable because both AI models independently developed their own encoding and decoding strategies without prior training or instructions. This autonomy marks a new milestone in AI capabilities.

This experiment highlights the increasing complexity and independence of artificial intelligence, which could transform how information security is approached.

Brian Chan

The ability of AI to autonomously create and decipher hidden messages points to both potential risks and opportunities across various fields, from data protection to creative innovation.