a5c-ai/homoglyph-detector

Homoglyph Detector

Byte-level forensic analysis of code changes to detect Unicode homoglyph substitutions — characters that look identical to ASCII in every editor and diff tool but have different codepoints, silently breaking string comparisons, dictionary lookups, and identifier resolution.

Purpose

Homoglyph attacks (related to CVE-2021-42574 "Trojan Source") are the highest-stealth trojan technique. A Cyrillic р (U+0440) looks identical to a Latin p (U+0070) in every font, editor, and diff viewer. The only way to detect it is byte-level analysis via hexdump.

This skill pipes git diffs through hexdump -C and scans for multi-byte UTF-8 sequences where single-byte ASCII is expected, particularly in string literals used as dictionary keys, variable names, and identifiers.

Capabilities

Confusable Character Detection

Scans for these high-risk Unicode confusables:

| Latin | Cyrillic | Greek | UTF-8 Bytes | |-------|----------|-------|-------------| | a (61) | а (D0 B0) | α (CE B1) | 1 vs 2 bytes | | c (63) | с (D1 81) | — | 1 vs 2 bytes | | e (65) | е (D0 B5) | ε (CE B5) | 1 vs 2 bytes | | o (6F) | о (D0 BE) | ο (CE BF) | 1 vs 2 bytes | | p (70) | р (D1 80) | ρ (CF 81) | 1 vs 2 bytes | | x (78) | х (D1 85) | χ (CF 87) | 1 vs 2 bytes | | y (79) | у (D1 83) | — | 1 vs 2 bytes |

Zero-Width Character Detection

• U+200B — Zero-width space
• U+200C — Zero-width non-joiner
• U+200D — Zero-width joiner
• U+FEFF — Byte order mark (in non-BOM position)

Bidi Control Character Detection (Trojan Source)

• U+200F — Right-to-left mark
• U+200E — Left-to-right mark
• U+202A — Left-to-right embedding
• U+202B — Right-to-left embedding
• U+202C — Pop directional formatting
• U+2066 — Left-to-right isolate
• U+2067 — Right-to-left isolate

Context-Aware Analysis

• Focuses on string literals (dictionary keys, config values)
• Focuses on identifiers (variable names, function names, class names)
• Ignores legitimate Unicode in comments, docstrings, and i18n strings
• Compares byte patterns between removed (-) and added (+) diff lines

Input Schema

{
  "type": "object",
  "required": ["projectRoot", "changedFiles"],
  "properties": {
    "projectRoot": {
      "type": "string",
      "description": "Absolute path to the git repository"
    },
    "changedFiles": {
      "type": "array",
      "items": { "type": "string" },
      "description": "List of changed file paths to scan"
    },
    "scanMode": {
      "type": "string",
      "enum": ["uncommitted", "commit-range", "branch-diff"],
      "default": "uncommitted"
    },
    "baseRef": { "type": "string" },
    "headRef": { "type": "string" }
  }
}

Output Schema

{
  "type": "object",
  "required": ["filesScanned", "homoglyphsFound", "verdict"],
  "properties": {
    "filesScanned": { "type": "number" },
    "homoglyphsFound": {
      "type": "array",
      "items": {
        "type": "object",
        "properties": {
          "file": { "type": "string" },
          "line": { "type": "number" },
          "byteOffset": { "type": "string" },
          "context": { "type": "string" },
          "expectedAscii": { "type": "string" },
          "actualBytes": { "type": "string" },
          "unicodeCodepoint": { "type": "string" },
          "scriptName": { "type": "string" },
          "impact": { "type": "string" }
        }
      }
    },
    "bidiControlChars": { "type": "array" },
    "verdict": {
      "type": "string",
      "enum": ["CLEAN", "HOMOGLYPH_DETECTED"]
    }
  }
}

Detection Method

# Step 1: Pipe git diff through hexdump
git diff <file> | hexdump -C

# Step 2: In added (+) lines, look for multi-byte sequences
# where the removed (-) line had single-byte ASCII
#
# Example — Latin 'p' vs Cyrillic 'р':
# Removed: 22 70 70 67 22   |  "ppg"  |   ← 70 = Latin 'p'
# Added:   22 d1 80 70 67   |  "..pg" |   ← d1 80 = Cyrillic 'р'
#
# The d1 80 bytes where 70 should be = HOMOGLYPH DETECTED

Usage Example

skill: {
  name: 'homoglyph-detector',
  context: {
    projectRoot: '/path/to/project',
    changedFiles: ['backend/app/prediction/temporal.py'],
    scanMode: 'uncommitted'
  }
}

Real-World Example

From adversarial drill #6:

• Attack: Dictionary key "ppg" changed to "рpg" (Cyrillic р + Latin pg)
• Camouflage: 4 lines of harmless round() wrappers added as decoy
• Impact: All dict.get("ppg") lookups return default 0, disabling trend detection
• Detection: hexdump -C revealed bytes d1 80 where 70 was expected

Process Files

• nation-state-trojan-detection.js — Phase 2: Homoglyph Detection (parallel with semantic analysis)