homoglyph unicode injection firewall

The problem

Standard security scanners and keyword filters operate on string literals. They cannot detect visually identical characters that use different codepoints or invisible control characters that reorder logic while appearing benign to the human eye.

What it does

• Scans text codepoint by codepoint to identify hidden bidirectional (bidi) control characters used in Trojan Source attacks.
• Detects cross-script homoglyphs where non-Latin characters (like Cyrillic or Greek) are substituted for Latin lookalikes in identifiers or domains.
• Identifies zero-width characters and invisible joiners that silently alter tokenization or bypass exact-match security checks.
• Flags normalization (NFC/NFKC) divergence where a document's rendered appearance differs from its underlying codepoint sequence.

Why this beats prompting it yourself

Most LLMs naturally interpret text as rendered, often missing the very byte-level deceptions this skill is designed to find. This tool follows a strict non-embedding rule, ensuring it describes threats via codepoint notation rather than inadvertently re-emitting dangerous bidi payloads in your environment.

Use cases

• Auditing Pull Requests for Trojan Source (CVE-2021-42574) code smuggling.
• Verifying that domain names or usernames are not using confusable character substitutions for phishing.
• Checking third-party agent skill files for hidden Unicode instructions before deployment.
• Comparing rendered text against actual codepoints to find invisible characters in security-sensitive keywords.

Known limitations

Uses a heuristic pass based on established confusable pairs. It is not a live query against the latest Unicode Consortium data and should not replace a dedicated Unicode-aware security library.