Homoglyph Detection

PromptShield detects mixed-script homoglyph attacks, where visually similar characters from different Unicode scripts are combined to create deceptive identifiers or instructions.

These attacks are commonly used in:

• prompt injection
• identity spoofing
• configuration manipulation
• phishing-style prompt attacks
• code review bypass

Why this matters

Humans read glyphs visually.

Computers interpret Unicode code points.

Two characters that look identical can be completely different:

a (Latin)     U+0061
а (Cyrillic)  U+0430

This enables spoofing like:

admin
admіn   ← Cyrillic "і"

They look identical in most editors.

They are not the same string.

This breaks validation, allow-lists, and policy checks.

Detection model

The PromptShield homoglyph detector:

• scans text for word spans
• inspects Unicode script composition per word
• detects suspicious Latin + Cyrillic or Latin + Greek mixing
• emits one diagnostic per word

The detector intentionally avoids flagging multilingual text to reduce false positives.

Rule

PSH001

Severity: CRITICAL

A word contains characters from multiple Unicode scripts that can be used for spoofing.

Example:

pаypal

The second character is Cyrillic:

p + Cyrillic "а" + ypal

Another example:

admіn

Where:

i → Cyrillic "і"

These words appear normal to humans but differ at the code-point level.

Suggested remediation

Replace homoglyph characters with characters from the intended script.

For identifiers, prompts, and configuration values:

• use ASCII when possible
• avoid mixed-script identifiers
• normalize input before validation

Design notes

PromptShield intentionally detects mixed-script composition, not individual confusable characters.

This avoids false positives in:

• multilingual documentation
• international content
• natural language text

Detection focuses on security-relevant misuse, not typography.

Mental model

Homoglyph detection protects against:

• identifier spoofing
• prompt impersonation
• policy bypass using confusable characters

It is conceptually similar to:

• IDN homograph protections in browsers
• Unicode spoofing detection
• authentication identifier validation