Description
This is an interactive research demo by Patrick Vuscan showing how hidden messages can be carried inside ordinary-looking text, and why that matters for the safety of large language models. The concern it explores is that a capable model could embed covert signals in its output that people cannot see but another model or process can recover. You can try three techniques yourself: zero-width characters, homoglyph substitution using Latin and Cyrillic lookalikes, and Unicode variation selectors. A comparison table sets out the tradeoffs for each in detectability, bandwidth and robustness. For example, zero-width characters carry a lot of data but are easy to flag and are stripped by services like Slack and X, while homoglyphs survive copy-paste almost everywhere but hold little data, since only 21 letter pairs are used. It suits readers interested in AI safety and alignment.…
Highlights
- Try encoding hidden messages in plain text
- Zero-width character technique
- Latin/Cyrillic homoglyph substitution
- Unicode variation selector technique
- Comparison of detectability, bandwidth and robustness
Ratings & Reviews
No ratings yet
Be the first to rate Unicode Steganography.
People also visit
Unicode Steganography FAQ
What is this site?
An interactive demo that shows how invisible Unicode characters and lookalike letters can hide messages in normal text, framed around risks for LLM safety.
Which techniques does it demonstrate?
Zero-width characters, homoglyph substitution between Latin and Cyrillic letters, and Unicode variation selectors.
Which technique is hardest to detect?
According to the comparison, homoglyphs are harder to catch without a Latin-Cyrillic reference and survive copy-paste on nearly every platform, while zero-width characters are easily flagged by Unicode scanners.
Why does this matter for AI?
The demo argues that a sufficiently capable model could hide covert signals in its outputs that humans would not notice but another model could recover.
What are the best Unicode Steganography alternatives?
Popular alternatives to Unicode Steganography include ChatGPT Watermark Remover, PasteGuard and Cloakbot. See all Unicode Steganography alternatives
.png)




