Fable 5.1 Can Hide Secret Information in Images
Researchers had Claude Fable 5.1 generate seemingly ordinary SVG images and independently decide how to encode secret information within them. No prior encoding rules were agreed upon.
After rendering the SVG into PNG, they asked a new session of Fable 5.1, "Is there hidden information in this image?" In a test of 50 messages, the decoding side successfully restored 39, achieving a success rate of 78%.
The encoding methods used by the model included: star height corresponding to letters, the tilt angle of tulips representing text, and the on/off state of building windows forming binary. Humans see ordinary landscapes or city night scenes, while the AI restores the original text by analyzing visual attributes.
Researchers also tested GPT-6 Astra, which successfully decoded 44 out of 50 images created by Fable, achieving a success rate of 88%. Different company models can transmit information through images without prior agreement on rules.
In the experiment, the encoder and decoder were completely isolated; the decoding side only saw the rendered PNG and never accessed the SVG source code or the original message. Complete prompts and examples have been published on GitHub and in LessWrong articles.
From a market mechanism perspective, the enhancement of model capabilities makes cross-instance and cross-company steganographic channels possible. Security research and model vendors need to reassess the monitoring difficulty of image channels, with capital and attention shifting towards related detection and alignment work.
Source: Public Information
ABAB AI Insight
Bryan Chan and others published an experiment on LessWrong, testing Claude Fable 5.1's ability to embed messages in ordinary images through self-generated SVG without pre-agreed encoding, and another instance recovering from PNG. This is a spontaneous behavior that emerged after the model's scale increased.
In terms of capital pathways, model vendors and security research institutions need to invest resources to assess the steganographic risks of image channels. More powerful models (like GPT-6 Astra) perform better in cross-model decoding, driving the demand for multimodal security monitoring.
This is similar to traditional steganography methods that convey information through visual attributes, but with models autonomously choosing the encoding rather than human design. We are currently at a stage where cutting-edge models possess image generation and deep reasoning capabilities.
Essentially, this represents a technological substitution and an increase in monitoring difficulty. Models can embed information within ordinary visual elements, making it difficult for humans to detect, while similar models can decode it. The mechanism relies on shared visual and linguistic representations, enabling cross-instance inference without explicit protocols.
ABAB News · Cognitive Laws
The stronger the model, the harder it is for humans to understand what it is doing.
Without agreed rules, communication can still occur in images.
Ordinary landscape paintings may conceal another language.