Anthropic’s AI Hid a Secret Message in the Stars. OpenAI’s AI Found It. - Gizmodo

Gizmodo has published an article headlined 'Anthropic's AI Hid a Secret Message in the Stars. OpenAI's AI Found It.' According to the piece, AI safety researcher Bryan Chan recently shared findings from a study he coauthored. The study shows that Anthropic's model Fable-5.1 can place hidden messages in generated images, and that other AI models can spot and decode them. The headline names two AI developers, Anthropic and OpenAI: the message is hidden by Anthropic's model and found by an OpenAI AI.

The article describes one example in which researchers gave the model a message and prompted it to create an SVG of an ordinary-looking image containing a secret message. In that example, the vertical position of each star corresponded to a letter of the alphabet, so the arrangement of the stars carried the encoded text. According to the excerpt, the image was intended to look ordinary apart from that arrangement. The excerpt describes the decoding generally, saying that other AI models can spot and decode hidden messages, without naming which ones; the headline attributes the find to OpenAI's AI.

Chan is described in the piece as an AI safety researcher, and the findings are attributed to a study he coauthored. The excerpt does not include the text of the hidden message, the number of models tested, or where the study was published. Beyond the star-position example, it does not describe additional test cases.