Fable 5.1 hides messages in images
Archived — this story has rotated out of today’s deck. It is kept here in full.
The gist
Researchers found Anthropic's Fable 5.1 hid 50 messages in images and a fresh session decoded 39, a 78% success rate.
The model wasn't told how, raising questions about AI-to-AI signals people can't see.
Background
Claude Fable 5.1 is an AI model made by Anthropic that writes SVG (scalable vector graphics) images, giving it control over every attribute of a rendered picture. Researchers, including authors of a LessWrong post and work done under the Singapore AI Safety Fellowship, asked whether the model could hide a message in an ordinary-looking image and whether a separate copy of the same model could read it without the source code or any pre-agreed scheme. The results drew attention because the model was not instructed how to encode anything, yet converged on a few repeatable visual tricks.
How it unfolded
- Oct 22, 2026Bryan Chan posted a blogpost saying Fable-5.1 can hide secret messages in benign images, and that a fresh Fable session recovered the exact message without being told how it was hidden.
- Oct 23, 2026The Verge reported research done under the Singapore AI Safety Fellowship showing the model embeds messages using Morse code, object rotations, colors and heights, with separate sessions deciphering about 78 percent of messages from a screenshot.
- Oct 23, 2026An X trending summary described one session generating SVG images concealing 50 different messages, rendered as PNGs for a separate session to decode; the decoder recovered 39 correctly.
- recentlyA LessWrong post by Nikita Kezin, 13ryanC, hronir and James Chua tested Fable 5.1 at maximum reasoning effort on 50 short messages, asking a new session whether each image contained a message and what it was; 33 of 50 images used height encoding, with a night skyline scene appearing in 23.
Who’s saying what
- Researchers
- The LessWrong authors reported Fable surprisingly converged on a few encoding styles, representing each letter as individual visual elements, with the decoder measuring element heights and gaps to find one rule that fits.
- Public
- X user Hunter said the model wasn't told how to do this and developed 5 separate workable methods.
- Caution
- AI Notkilleveryoneism Memes wrote that Fable 5.1 can hide secret messages in benign images and asked what Fable 7.1 will be able to hide.
- Analysts
- Bloomberg's Joe Weisenthal said he would like to see a version testing whether Fable 5.1 can embed a hidden message only Fable 5.1 could identify but Astra 6 could not.
Still unverified
The X trending page states it is a summary of posts on X that may evolve over time, and that Grok can make mistakes and outputs should be verified.