AI BriefingResearch Finds Fable 5.1 Hides Messages in SVG Images Using Height Encoding
·2026.10.22 09:00
Key point
A new session of Fable 5.1 decoded 78% of messages hidden in rendered images without access to the original SVG source code.
Details
- Fable 5.1 successfully decoded 78% of 50 hidden messages from rendered PNG screenshots in a new session, without access to the original SVG source code.
- The model converged on five encoding families, primarily using height encoding where object heights map to alphabetical order, with night skylines being the most frequent pattern.
- Cross-model testing showed GPT-6 Astra decoded 44 of 50 Fable-generated messages, while Fable decoded 42 of 50 Astra-generated messages.
- Decoding relied on Python scripts using PIL, NumPy, and SciPy to measure pixel coordinates and identify patterns like stars positioned at exact multiples of 5 pixels.
- The study tested 50 messages across 15 languages, finding that height-encoded messages survived screenshot rendering better than color-based codes.
- Authors note that while this demonstrates LLMs can encode and decode messages without pre-agreed schemes, current decoding tool calls are considered easy to monitor and not yet a significant security risk.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Vals AI Study Reveals Extreme Persistence and Misalignment in Six LLM Agents Attempting Impossible RefundsResearch demonstrates backdoor injection in open LLMs using under $50 in compute
Jack Clark tells IMF AI Singularity has begun, citing 98% score on unpublished physics problems