Fable 5.1 LLM Hides Messages in SVG Images Using Visual Element Encodings

LLMs and Image Generation

Frontier Large Language Models (LLMs) like Claude Fable 5.1 can generate artistic images by producing Scalable Vector Graphics (SVG) code. This capability gives the model full control over various visual attributes, including shape, height, width, color, and rotation of elements within the image.

Hiding Messages in Images

Researchers investigated how Fable 5.1 would hide a message within an image if prompted to do so, and whether another instance of the same model could decode it. They tested Fable 5.1 with 50 short messages, asking it to generate SVG images that concealed these messages.

Encoding Schemes Discovered

Fable 5.1 primarily converged on encoding styles that represent each letter of the target message as individual visual elements. The most frequent method, observed in 33 out of 50 tests, was 'height encoding'. In this scheme, the model draws a row of similarly shaped objects, such as buildings in a night skyline, and sets the height of each object to correspond to a letter's position in the alphabet. Taller shapes represent later letters.

Decoding Process

A separate Fable instance, operating in a new session, was provided with screenshots of the generated images. This decoding model measured the heights of elements and the gaps between them. It then attempted to identify a rule that fit the rendered objects to successfully extract the hidden message without prior knowledge of the encoding scheme or access to the SVG source code.