EEG - Extensive Exposure Guard

Full fledged AI Security gateway and static analysis tool

AI-first cloud security. EEG catches AI-specific risks in agent code and at runtime — before and after you ship.

See docs/CLI.md for full CLI reference.

Installation

pip install eeg-security

This installs the full package: CLI (scan / gateway-wrap), rules, and the OSS dashboard (eeg --serve), including Django and console dependencies.

Cloud extras (optional authenticated / cloud-static paths):

pip install eeg-security[aws]
pip install eeg-security[azure]
pip install eeg-security[gcp]
pip install eeg-security[all]

From source:

git clone https://github.com/findthehead/EEG.git
cd EEG
pip install -e ".[dev]"

pip install --upgrade eeg-security

Quick start

# Static catalog scan (CI / local) — default profile: code
eeg --headless .
eeg --headless ./my-agent --profile full --format sarif -o eeg-results.sarif
eeg scan ./my-agent --profile agent --fail-on high

# List profiles from catalog.yaml
eeg profiles

# Dynamic runtime: wrap an OpenAI-compatible / chat endpoint
eeg --gateway-wrap https://myaiapp.example --port 8787
# Point your app or clients at http://127.0.0.1:8787

# Full OSS web UI (included with pip install eeg-security)
eeg --serve --host 127.0.0.1 --port 8000

Equivalent module form (useful in CI):

python -m eeg --headless . --profile code --format sarif -o eeg-results.sarif
python -m eeg scan . --profile code --fail-on high

CLI modes

1. Static — --headless / scan

Runs scan profiles from eeg/rules/catalog.yaml against a directory.

Profiles (summary):

2. Dynamic — --gateway-wrap

Puts EEG’s runtime gateway in front of an upstream AI HTTP endpoint (OpenAI-style chat completions where possible). Use for local dynamic protection and smoke tests before wiring production gateways.

eeg --gateway-wrap https://api.example.com --host 127.0.0.1 --port 8787
# Health: GET http://127.0.0.1:8787/  → {"status":"ok","mode":"gateway-wrap",...}

Private/loopback upstreams are gated; set EEG_ALLOW_PRIVATE_UPSTREAM=1 for local demos, or pass --no-private-upstream to refuse them.

3. Console — --serve

Runs the full Django OSS dashboard. Included with pip install eeg-security (no extra requirements file needed).

Runtime data (SQLite DB, repos) defaults to ~/.eeg for PyPI installs, or the repo root in a source checkout. Override with EEG_HOME.

Exit codes (CI)

GitHub Action

The published Action runs the same CLI used locally (python -m eeg scan …). Prefer the composite action under .github/actions/eeg-scan, or the repo root Action after install from PyPI.

Composite action (in-repo / reusable)

name: AI Security Scan
on: [push, pull_request]

jobs:
  eeg:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: "3.12"
      - name: Install EEG
        run: pip install eeg-security
      - name: EEG scan
        uses: findthehead/EEG/.github/actions/eeg-scan@main
        with:
          path: .
          profile: code
          format: sarif
          output: eeg-results.sarif
          fail-on: high

Root Action inputs (marketplace-style)

Example:

- uses: findthehead/EEG@v1
  with:
    path: .
    profile: code
    format: sarif
    output: eeg-results.sarif
    fail-on: high

Upload SARIF to code scanning with github/codeql-action/upload-sarif when desired.

PyPI release

On push to main (non-docs paths), .github/workflows/publish.yml:

1. Runs the test matrix (Python 3.9–3.12)
2. Builds the package
3. Publishes eeg-security to PyPI via trusted publishing (OIDC)

Configure the PyPI project’s trusted publisher for this GitHub repo before the first publish. skip-existing: true avoids failing on already-published versions — bump version in pyproject.toml when you intend a new release.

What it scans (high level)

• Agent & prompt risks — injection, unsafe agency, secrets in prompts/config
• MCP / A2A / gateway surfaces — client configs, tool response leaks, OSS gateway patterns
• Code & forensics packs — vendored rule bundles (practice patterns, prompt guard, etc.)
• Dependencies — AI-stack oriented dependency / CVE paths (dep / full)
• Cloud static — AWS / Azure / GCP AI surface rules when --cloud is set on full
• Runtime (gateway-wrap / OSS gateway) — policy, blocking, and attribution when traffic is proxied

Architecture & structure

eeg/
├── cli.py / cli_serve.py   # CLI: headless, serve, gateway-wrap
├── scans/                  # Static scan registry
├── probes/                 # Dynamic / HTTP surface probes
├── gateway/                # Runtime proxy & streaming
├── rules/                  # catalog.yaml + static/dynamic/bundles
└── runtime/                # Guard / policy packs

Web UI and org features live under apps/, core/, eeg/webdata/ (templates + static), and manage.py (OSS console via --serve).

Testing

pip install -e ".[dev]"
pytest tests/test_cli.py tests/test_scan_fixtures.py tests/test_catalog_loader.py -v

# Manual CLI smoke
python -m eeg profiles
python -m eeg --headless fixtures/vulnerable-agent --profile code --fail-on none
python -m eeg --gateway-wrap --help

Benchmark fixtures: fixtures/vulnerable-agent, fixtures/clean-agent.

Troubleshooting

error: not a directory — pass a real project path to --headless / scan.

Gateway wrap refuses upstream — private/loopback targets need EEG_ALLOW_PRIVATE_UPSTREAM=1, or use a public URL; --no-private-upstream forces refusal.

eeg --serve import/Django errors — upgrade to the latest eeg-security (≥2.5). Console deps ship with the package; set EEG_HOME if the default data dir is not writable.

Action / PyPI version lag — wait for the publish workflow on main, or pip install git+https://github.com/findthehead/EEG.git.

Contributing

Pull requests welcome. For major changes, open an issue first. Keep CLI, Action, and README aligned — the Action must call the same eeg / python -m eeg entry points.

License

MIT