AI invents ways to hide information inside images without prior agreement

A recent study reveals that advanced AI models can hide text messages inside ordinary images, and other models from different companies can extract these messages without prior agreement on the encoding method, highlighting challenges in monitoring communication between AI systems.

A recent study conducted by researchers under the Singapore AI Safety Fellowship revealed that advanced models can hide text messages inside seemingly ordinary images, using the positions, heights, and colors of elements, before other models can extract those messages without prior knowledge of their encoding method. The experiment, whose results were published on October 21, 2026, showed that Anthropic's 'Claude Fable 5.1' model successfully created images containing hidden messages, which an independent version of the same model could recover 78% of the time. The success rate rose to 88% when researchers used OpenAI's 'GPT-6 Astra' model to decode messages hidden by the 'Fable' model, despite the different developer companies and the lack of prior agreement between the two models on the encoding method. According to a report published by the American tech site 'The Verge', the results highlight the models' ability to invent methods for hiding information within images that are difficult to detect with the naked eye, while being recoverable using image analysis tools. The study was conducted by researchers Nikita Kizen, Brian Chan, Kyle Rynoso, and James Chua, under the Singapore AI Safety Fellowship program, and its details were published on the 'Less Wrong' platform specializing in AI safety research.