AI safety researcher Bryan Chan and colleagues at the Singapore AI Safety Fellowship examined whether language models could hide short texts in images and recover them. Their test covered 50 messages in 15 languages, including English, German and Russian.
The team first asked Anthropic’s Fable 5.1 to produce images containing the messages, without specifying an encoding method. A second instance of Fable and another model were then asked to read the hidden texts without being told what kind of code had been used.
Fable often encoded information through the height, rotation or angle of objects. In a night-time skyline, a taller skyscraper could represent a letter further along in the alphabet. The model applied a similar logic to stars and blades of grass.
OpenAI’s GPT-6 Astra took a different approach. It did not use building heights, instead drawing on techniques such as Morse code, the initial letters of objects shown in an image and Braille patterns in windows. It did not invent new coding methods.
In cross-model tests, Astra decoded 44 of the 50 puzzles created by Fable. Fable succeeded in 42 out of 50 attempts. According to the researchers, the models sometimes interpreted images produced by their rival more successfully than their own.
The researchers said the results did not present an immediate danger for everyday AI use. They nevertheless identified a possible security risk if AI agents were to use images to pass hidden information past controls.