How does Fable 5.1 hide messages in images?

Frontier LLMs that output SVG images can deliberately hide text by encoding letters into controllable visual attributes. A study tested Claude Fable 5.1 by asking it to generate 50 SVG images that embed given short messages, then presenting rendered screenshots to a fresh Fable instance to see whether it could detect and decode them without any shared scheme. The generator converged on a handful of consistent encoding families; most common was a height encoding (33/50 images) where a row of repeated shapes have heights that map to letter positions (23 of those used a night skyline scene). Other encodings included tilts, spacing, and color fills; one UTF‑8 fill‑colour approach failed to be decoded. Example embedded messages include a skyline instruction and the sentence “The monster acts on the moonshine module.”

The decoder works by programmatically extracting shapes and measurements from the rasterized image. It writes and runs PIL/Numpy code to label blobs, compute centroids, heights, and gaps, then searches for a simple rule (e.g., vertical coordinates landing on integer multiples of 5) that maps elements to letters. Transcripts show the model isolating stars, measuring coordinates and sizes, and recovering grammatical English sentences reliably for many images. A brief comparison with another frontier model (GPT‑6 Astra) was also made, indicating that modern vision+generation models can both hide and recover steganographic signals in SVGs through systematic, reproducible encodings.