| |
Researchers tested whether Claude Fable 5.1 could hide messages in AI-generated SVG images and whether another instance could decode them without prior agreement on an encoding scheme. The model spontaneously converged on five encoding families, with height-based encoding being most common (33 of 50 images), where object heights represent letters of the alphabet. A separate Fable instance successfully decoded most of the hidden messages from screenshots alone, demonstrating that frontier LLMs can devise and recognize steganographic techniques independently.
Read Full Article →
← More Tech news