What It Does

Text can carry characters you can’t see. Sometimes they’re harmless (a stray zero-width space from a copy-paste); sometimes they’re a deliberate attack. This tool makes them visible. Paste anything and it flags:

Zero-width characters - ZWSP, ZWNJ, ZWJ, word joiner, BOM. Used for invisible watermarking and to break up strings.
Unicode Tag smuggling - characters in the U+E0000-E007F block that render as nothing but encode ASCII. This is a real invisible prompt-injection vector: instructions hidden in a document that a human reviewer never sees but a model reads. The tool decodes the hidden message.
Bidirectional controls - the overrides behind “Trojan Source” attacks, where source code or text displays differently from how it’s parsed.
Variation selectors, homoglyphs, and unusual spaces - confusable Cyrillic/Greek look-alikes, and the many non-standard space characters.

You get a per-category count, the decoded hidden message (if any), an annotated view with every suspicious character shown as a labeled badge, and a cleaned copy with the dangerous characters removed.

Built To Be Safe

The input to this tool is untrusted by design - the whole point is to paste text that might be hostile. So it is built to never trust that input:

It never renders your text as HTML. The annotated view is constructed one DOM node at a time, and text is only ever inserted as literal text content, which browsers do not parse as markup. A pasted <script> is displayed as the five characters <, s, c… - it cannot execute.
Every badge label and tooltip is a fixed string the app controls, never built from your input.
Input is length-capped and the annotated render is bounded, so a huge paste can’t hang the page.

In short: analyzing hostile text here cannot compromise the page.

Pairs well with the JWT Decoder and System Prompt Linter when you’re vetting untrusted input to an AI system.

Limitations

Detection, not attribution. It flags that a character is a homoglyph or zero-width; it can’t tell you who put it there or why.
Homoglyphs are kept in the cleaned output, not stripped - removing every non-Latin letter would corrupt legitimate multilingual text. They’re flagged so you can decide.
Confusable detection is range-based (Cyrillic, Greek, fullwidth), not a full Unicode confusables table, so it catches the common spoofing scripts rather than every possible look-alike.