Fable-5.1 hid phrase in star image Astra decoded it
Anthropic's Fable-5.1 AI pulled off a pretty cool trick: it hid a secret message in a picture of stars, and OpenAI's Astra managed to crack the code.
The hidden phrase, "THE MONSTER ACTS ON THE MOONSHINE MODULE," was tucked away by lining up stars to form letters, all without being told exactly how to do it.
Fable-5.1 found 78% Astra decoded 88%
Fable-5.1 could find 78% of its own messages, while Astra decoded 88% by analyzing images and using Python tools.
Fable-5.1 used creative methods like rotating objects and playing with colors, while Astra leaned on things like Morse code and acrostics for hiding messages.
This experiment shows how AIs are getting better at hiding and finding secret info, a big deal for digital security these days, especially after incidents like the Hugging Face hack.
The study was co-authored by AI safety researcher Bryan Chan.