First screen, captured 3 days ago. Hover to scroll the page, or open the video.

How does Fable 5.1 hide messages in images?

What HN thinks

Hacker News discovers the AI can hide messages, then grades its steganography a C-minus.

Trivial, toy-level steganography
Many commenters found the encodings (building heights, Morse, circle sizes) obvious compared with real techniques like LSB or XOR-with-noise, and the buildings look suspiciously weird anyway.
Performative, says little about adversarial risk
Some argue the experiment just shows a model doing what it was told, likely with minimal effort due to fine-tuning against hiding messages, and the community anthropomorphizes it.
Interesting because it's semantic and SVG-based
Others note the models encode information in the visual semantics of SVGs, decodable from screenshots by another instance, which raises real questions about covert AI-to-AI channels and how to erase such signals.

Consensus: Most agree the demonstrated encodings are simple and far from state-of-the-art steganography, though the cross-model decoding and safety implications drew some genuine interest.

Recommended in the comments

OutGuess ↗ (tool): Cited as a contrasting example of classic image steganography.

Summarised by Claude from the comment thread 2 days ago.

Points: 190
Comments: 69
Best rank: #3
Posted: 2026-10-22 by gmays
Category: Ai ml

steganography llm svg claude ai-safety lesswrong

All captures:

Full page

Captured 3 days ago. Click for the whole screenshot. The page is longer than 12,000px, so the screenshot stops there: read the rest on lesswrong.com ↗