The researchers used the model at maximum reasoning effort and generated one image for each of 50 messages.
A separate Fable 5.1 session then received screenshots of those images, without the SVG source and without a previously agreed encoding scheme. The authors asked it whether an image contained a message and, if so, to recover it. They describe five families of visual encodings. In 33 of the 50 images, the model used height: repeated objects such as buildings represented letters by their relative height, with gaps separating words.
The post says the decoding model inspected element heights and spacing to infer a rule that fit the image. It also briefly compares the task with GPT-6 Astra, but the supplied account gives no comparable aggregate result for that model.
This is a small, source-described experiment rather than evidence that hidden-image messaging is prevalent in deployed systems. It does not establish the reliability of the technique across models, its ability to evade any particular moderation or security control, or a real-world incidence rate.