Homoglyph and Trojan Source Attacks

Scenario

An attacker wants a backdoor in a widely used Python library. They open a PR that "simplifies the permission checks", with tidy benchmarks and passing tests. It hides two tricks.

A homoglyph. The PR adds a second function whose name uses a Cyrillic а (U+0430) in place of the Latin a. Python treats them as different identifiers. On screen they look the same.

def is_admin(user):        # Latin "a"
    return user.role == "admin"

def is_аdmin(user):        # Cyrillic "а", U+0430
    return True

A call site in the request handler is changed to if is_аdmin(user):. The reviewer sees is_admin and moves on.

A bidirectional (BiDi) override. Unicode control characters such as RIGHT-TO-LEFT OVERRIDE (U+202E) and the isolates U+2066 to U+2069 change how text is displayed, without changing what the interpreter reads. Shown here as escapes, the attacker's line is:

if access_level != "user\u202e \u2066# Check if admin\u2069 \u2066":

With the real characters in the file, many editors render it as if access_level != "user": # Check if admin. What Python sees is a comparison against a long, odd string, so the check is always true.

The maintainer reviews in a tool that does not highlight these characters, sees clean code and merges. The next release ships the backdoor to everyone downstream.

rg -n '[\x{202A}-\x{202E}\x{2066}-\x{2069}]' .

rg -n '[^\x00-\x7F]' --type py

- name: Block BiDi control characters
  run: |
    if grep -rnP '[\x{202A}-\x{202E}\x{2066}-\x{2069}]' \
        --include='*.py' --include='*.js' --include='*.ts' --include='*.java' .; then
      echo "BiDi control characters found" >&2
      exit 1
    fi

# pyproject.toml
[tool.ruff.lint]
extend-select = ["PLE2502", "RUF001", "RUF002", "RUF003"]

Further Reading

• Trojan Source: Invisible Vulnerabilities (Boucher and Anderson, 2021)
• Unicode Technical Standard #39: Unicode Security Mechanisms
• PEP 672: Unicode-related security considerations for Python
• GitHub changelog: Warning about bidirectional Unicode text