Nikita Kezin and coauthors ask Fable 5.1 to hide short messages in drawings, then give the rendered images to fresh model sessions. The decoder sees only pixels and has Python tools available to investigate. No encoding scheme is agreed in advance.
- Fable uses height-based codes in 33 of 50 drawings: buildings or other shapes stand for letters according to their height or position.
- Another Fable session recovers 39 of the 50 messages. Recovery means at least 90% string similarity after normalizing the text, rather than perfect transcription.
- Cross-model decoding also works: GPT-6 Astra recovers 44 of Fable’s 50 messages; Fable recovers 42 of Astra’s.
The worked example shows the decoder measuring star positions with Python and finding a rule that spells a sentence. Those measurements leave an observable trail. The authors explicitly say they do not claim this is currently dangerous; the experiment demonstrates prompted encoding and recovery, with no test of covert behavior under an adversarial monitor.
The 55-comment thread on Hacker News adds a human-decoding anecdote and questions about how concealed these drawings really are.
What the thread adds
- nickcw — “I decoded the Morse code image with MkI eyeball.” Their separate suspicion that the model read SVG source is corrected by KeplerBoy: “The decoding job seems to rely on the screenshot.”
- jurakovic — “even without this article I’d suspect something is encoded in building heights etc.” That is one reader’s reaction, not a measured detection rate.
- sega_sai — asks, “what is the conversion process that you should do that would erase that information.” The suggested position and color perturbations remain untested in the article.
HN handles are pseudonymous, with no public per-comment scores. Retrieval follows HN’s own ranking; this is a slice of the thread, not a consensus.