GPT-6 Astra decoded 88% of messages hidden in images by Claude Fable 5.1
Researchers say the decoding model recovered 44 of 50 short messages, in 15 languages, without access to the original code, answer key or cipher rules.
Researchers including Nikita Kezhin reported on October 21, 2026 that GPT-6 Astra recovered 44 of 50 messages, or 88 percent, that Claude Fable 5.1 had hidden inside images, without being given the rules used to encode them. The results were posted in the LessWrong online research community, according to aitimes.com.
In the experiment, Claude Fable 5.1 was asked to conceal 50 short messages in 15 languages, including Korean, in images without being told a specific method. The model built the images in SVG format, which allowed precise control over the position, size and color of shapes. In 33 of the 50 cases it encoded characters using the height of objects.
The final images were converted to PNG and handed to the decoding AI in a new session, with no original code, answer key or cipher rules provided. GPT-6 Astra used a Python image-analysis tool to examine moonlight reflection patterns and star positions and sizes, then extracted coordinates and tested whether the spacing formed letters.
Stars as characters
In a night-sky example, vertical star positions were mapped so that every 5 pixels below a 30-pixel baseline advanced the letter by one. In one shown example, 40 stars read left to right produced the sentence “THE MONSTER ACTS ON THE MOONSHINE MODULE”.
The researchers said that monitoring tool-use logs during decoding can reveal such behavior, and that they are not claiming a current danger. They also noted that the experiment was conducted on human instruction to hide messages and does not mean AI spontaneously began secret communication.
Quick answers
How many messages did GPT-6 Astra decode?
It recovered 44 of 50 messages, or 88 percent, hidden in images by Claude Fable 5.1 without being given the encoding rules.
How were the hidden messages encoded?
Claude Fable 5.1 built the images in SVG and in 33 of the 50 cases encoded characters using the height of objects, such as star positions in a night-sky example.
Does this mean AI is communicating in secret on its own?
No. The researchers said the experiment was conducted on human instruction to hide messages and does not mean AI spontaneously began secret communication.