VigilML

Security scanner for the AI development lifecycle.

pip install vigilml && vigilml scan .

Requires Python 3.10+.

Contents

• What it catches
• Quick start
• All CLI options
• Suppressing findings
• CI/CD integration
• Real findings on real repos
• Available scanners
• Contributing
• Licence

What it catches

VigilML runs 7 scanners over your project and reports findings with a file, a line number, a severity, and a remediation. The table below lists specific detectors from each scanner, not a general summary.

Quick start

# Scan the current directory
vigilml scan .

# Scan with JSON output for CI/CD pipelines
vigilml scan . --json

# Run only specific scanners
vigilml scan . --scanners credentials,model_files

All CLI options

Suppressing findings

VigilML supports three suppression comments, checked directly in your source files. All three require an explicit comment — there is no way to silently disable a finding without leaving a trace in the code.

Inline — suppresses a single line. Use this for one isolated false positive, such as a test fixture value that happens to match a credential pattern.

# Known false positive: fixture value used only in tests
TEST_API_KEY = "sk-test-51H8xJ2KL9mN3pQrStUvWxYz12345"  # vigilml: ignore

Block — suppresses every line between the two markers. Use this for several consecutive lines that are all false positives, such as a block of demo credentials in a tutorial notebook.

# vigilml: ignore-start
# Demo credentials for the onboarding notebook. Never real, rotated
# before every workshop.
DEMO_HF_TOKEN = "hf_demoTokenNotARealSecret1234567890"
DEMO_OPENAI_KEY = "sk-demo-not-a-real-openai-key-000000000000"
# vigilml: ignore-end

File-level — suppresses every finding in the file. Use this only when an entire file exists to contain example patterns, such as a scanner's own test fixtures or its pattern definitions.

# vigilml: ignore-file
"""Fixtures for the credential scanner's unit tests.

Every string below is a synthetic pattern the scanner is meant to
detect, not a real secret.
"""

CI/CD integration

Basic version — fails the build on any finding, of any severity:

name: Security scan
on: [push, pull_request]
jobs:
  vigilml:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: "3.11"
      - run: pip install vigilml
      - run: vigilml scan . --no-colour

Strict version — narrows to the scanners whose findings are most often CRITICAL/HIGH (--scanners), and writes a config that raises every rule's min_severity to HIGH so the exit code reflects severity, not just presence:

name: Security scan (strict)
on: [push, pull_request]
jobs:
  vigilml-strict:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: "3.11"
      - run: pip install vigilml
      - name: Write a CRITICAL/HIGH-only config
        run: |
          cat > .vigilml-strict.yml << 'EOF'
          version: 1
          rules:
            credentials:
              min_severity: HIGH
            model_files:
              min_severity: HIGH
            dependencies:
              min_severity: HIGH
            prompt_injection:
              min_severity: HIGH
          EOF
      - run: >
          vigilml scan . --config .vigilml-strict.yml
          --scanners credentials,model_files,dependencies,prompt_injection
          --no-colour

Real findings on real repos

All repos scanned with vigilml scan . on unmodified public code.

Available scanners

Contributing

See CONTRIBUTING.md for setup, code standards, and how to report a false positive. Open an issue at github.com/sharmaamanrajesh/vigilml/issues before starting any large change.

Licence

MIT — see LICENSE.