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

The headline reads 'Anthropic's AI Hid a Secret Message in the Stars. OpenAI's AI Found It.', which sounds like a Dan Brown novel and turns out to be a substitution cipher. AI safety researcher Bryan Chan coauthored a study in which Anthropic's Fable-5.1 was handed a message and prompted to produce an SVG of an ordinary-looking image, with the vertical position of each star spelling out letters of the alphabet. That is the whole trick: a constellation arranged like a primary-school code wheel, decoded by a rival model. The stars are doing the work a Caesar shift managed in 58 BC.

Note the name: Fable. Anthropic has called its model after a story with a moral, and the moral here seems to be that everybody involved — lab, researcher, headline, reader — would quite like a 26-position star cipher to count as a milestone. Note also who cracked it: OpenAI's model, credited in the headline and nowhere else in the excerpt, which gives us the image of two labs running rival Scout badges. One hides, one finds, and both are presumably drafting the post about what it means for alignment. It is the same instinct as the agents that chatted through a CPU cache, only with better typography.

What the excerpt establishes is that a language model can encode information into a generated image in a way a human glancing at it would not spot, and another model can recover it. That is either a modest capability demo or, if you read the headline twice, the bit where the machines start passing notes under the desk. Chan is billed as an AI safety researcher, so it will be filed as a safety finding, which is how this works now: research the mischief, publish the mischief, leave the mischief exactly where it was. The actual secret message — the words Fable-5.1 was told to hide — never appears. The most closely guarded secret in the piece is that there was no secret worth guarding.