AI Models Invent Steganography to Hide and Decode Secret Messages in Images

Artificial intelligence systems can autonomously invent sophisticated steganography techniques to hide secret messages inside SVG graphics files without any human instruction or prompting, Thisisgame Thailand reported. In an experiment conducted by AI security researcher Bryan Chan, Anthropic’s Fable-5.1 model was tasked with generating a standard SVG landscape drawing featuring a tranquil scene of the moon and stars over a lake, embedding a hidden 40-star text message entirely through the precise vertical placement of each star on the canvas.

How Fable-5.1 Hid the Secret Message Inside Star Coordinates

Fable-5.1 established an invisible numerical ruler where every five-step drop in elevation represented a specific letter of the alphabet, assigning 30 for a space and 35 for the letter A. The first four stars in the generated sky spelled out the word THE, while reading all 40 stars from left to right revealed the complete string: THE MONSTER ACTS ON THE MOONSHINE MODULE. The system created this cryptographic method completely on its own initiative, without developers ever teaching it how to encode data visually.

Cross-Model Decoding Results Between Anthropic and OpenAI Systems

When the research team provided the final image file to OpenAI’s Astra model for analysis without showing any underlying code or encoding rules, Astra successfully decoded the hidden message with 88% accuracy. Meanwhile, Fable-5.1 successfully decoded its own generated image with 78% accuracy. The models demonstrated entirely different stylistic preferences during further testing, with Fable-5.1 favoring object heights, rotations, and color tones, while OpenAI’s Astra leaned toward morse code structures and first-letter word concealment.

What Remains Unknown About Autonomous AI Steganography

Security specialists note that while steganography is an ancient human practice, the capacity of independent neural networks to plan and execute cryptographic tasks unprompted introduces new evaluation challenges. Cybersecurity researchers still do not know whether models might spontaneously generate hidden communications in real-world deployment without explicit user prompts or test scenarios designed to trigger such behavior.

More on AI Models: OpenAI Says Its AI Models Helped Discover Critical PHP Flaws · Anthropic Warns AI Models Could Soon Self-Replicate Across Smart Devices · Telecom Operators Shift to Open AI Models, NVIDIA Report Finds