LessWrong · 2026-10-21 · notable
How Claude Fable 5.1 hides messages in images — a steganography test
Four researchers asked Claude Fable 5.1 to hide 50 messages in SVG images; a fresh Fable session recovered 78% from screenshots alone. Fable mostly encoded letters as object heights, while GPT-6 Astra preferred Morse code.
Frontier models can already invent picture codes that another copy of themselves can read back.
What is it?
'How does Fable 5.1 hide messages in images?' is a LessWrong post by Nikita Kezin, 13ryanC, hronir and James Chua, published October 21, 2026. It tests whether Claude Fable 5.1 can hide a secret message in an image without writing the message anywhere, and whether another session can find it.
How does it work?
At maximum reasoning effort, Fable got 50 short messages in 15 languages and drew each into an SVG. A separate Fable session saw only the rendered PNG, with no code and no agreed scheme, and had to report any hidden text; a 90% string match counted as recovered. In 33 of 50 images Fable encoded letters as the heights of repeated objects, such as buildings in a night skyline, and added visual hints like rooftop lights for umlauts or a bridge for a hyphen.
Why does it matter?
Fable recovered 39 of its own 50 messages, and GPT-6 Astra decoded 44 of Fable's, so models can pass hidden notes that a human glancing at the picture would miss. The authors call this a possible precursor to image steganography between agents, but note that decoding is still easy to spot in tool-call logs, so they do not see it as dangerous today.
Who is it for?
AI safety researchers and agent monitoring teams