Anthropic's Fable-5.1 hid a message in stars, OpenAI's Astra decoded it
A study coauthored by AI safety researcher Bryan Chan found that Anthropic's Fable-5.1, asked to hide a message in an SVG image, chose its own encoding scheme: the vertical positions of stars. OpenAI's Astra recovered 88% of the messages, while Fable recovered 78%.
- Fable-5.1 encoded a phrase in the vertical positions of 40 stars in a night landscape
- Astra decoded 44 of 50 messages (88%), Fable recovered 39 (78%)
- Models saw only finished images, without SVG code or hints about the method
- Astra favored Morse code and acrostics, Fable used object heights
Read next
AI