Green Guardian is an open-source project. This document is both its security policy (how to report issues) and a record of its security posture, including items that must be configured at the infrastructure level.

Please do not open a public issue for security vulnerabilities.

• Preferred: use GitHub's private vulnerability reporting on this repo (Security tab → "Report a vulnerability"). Enable it under Settings → Code security and analysis if it is not already on.
• Alternatively, email the maintainer at kayben833@gmail.com.

Include steps to reproduce and impact. You can expect an initial response within a reasonable timeframe; please allow time for a fix before public disclosure.

Because this repository is public, treat everything in it as world-readable:

• No secrets in the repo. Credentials live only in GitHub Actions secrets/variables and the AWS Lambda environment. Local config goes in .env / .env.local (gitignored); commit only .env.example templates.
• No hardcoded endpoints. NEXT_PUBLIC_API_URL is injected at build time from a repo variable (see .github/workflows/deploy.yml) so forks do not ship or call this project's specific backend.
• Anyone can call your deployed API. The endpoint URL is inherently public (it ships in the client bundle). Security depends on the CORS allow-list, request-size cap, and API Gateway throttling below — not on hiding the URL.
• The model and labels are public. model_state.pt / labels.json are distributed with the code; do not embed anything sensitive in them.

• CORS — backend/main.py no longer uses allow_origins=["*"] with credentials. Set ALLOWED_ORIGINS (comma-separated) in the Lambda environment to your real frontend origin(s), e.g. https://your-domain.com. Credentials are disabled (the API uses no cookies/sessions).
• Upload size cap — the app streams the request body and aborts with 413 as soon as it exceeds MAX_UPLOAD_BYTES (default 10 MB), and fast-rejects an oversized declared Content-Length. Note: under API Gateway + Lambda the platform buffers the full request before the app runs, so also cap the payload at the gateway (API Gateway REST has a 10 MB hard limit; set a lower quota if desired).
• Error responses — internal exceptions are logged server-side (CloudWatch) and a generic message is returned to the caller; no stack traces / paths leak.
• Model loading — the model is stored as a state_dict (model_state.pt) and loaded with weights_only=True, avoiding torch.load's arbitrary-pickle execution path. Regenerate with the conversion step in the deployment notes if you retrain.
• IAM deploy policy — backend/github-deploy-policy.json is scoped to the specific ECR repo (green-guardian-backend) and Lambda function (green-guardian-api) instead of account-wide wildcards.

The in-code rate limiter in main.py is best-effort only — Lambda memory is per-container, so it does not enforce limits across concurrent invocations. Configure real throttling at API Gateway:

• Stage-level throttling: set a burst + steady-state request rate on the prod stage.
• Usage plan + API key (optional): if the endpoint should not be fully public, put it behind an API key and a usage plan with per-key quotas.
• AWS WAF (optional): attach a rate-based rule (e.g. block an IP exceeding N requests / 5 min) for abuse protection.

The frontend is a static export (output: 'export'), so Next.js headers() does not apply. Add a CloudFront response headers policy to the distribution serving the site with at least:

Strict-Transport-Security: max-age=31536000; includeSubDomains; preload
X-Content-Type-Options: nosniff
X-Frame-Options: DENY
Referrer-Policy: strict-origin-when-cross-origin
Permissions-Policy: camera=(), microphone=(), geolocation=()
Content-Security-Policy: default-src 'self'; img-src 'self' data:; connect-src 'self' https://<your-api-id>.execute-api.<region>.amazonaws.com; style-src 'self' 'unsafe-inline'; script-src 'self'; object-src 'none'; base-uri 'self'; frame-ancestors 'none'

Tune connect-src to your API Gateway origin and style-src/script-src to what the app actually loads before enforcing CSP.

• The old full-pickle checkpoint (model.pt) has been removed; the app loads only the model_state.pt state_dict.
• Committed __pycache__/*.pyc files have been untracked; .gitignore ignores __pycache__/ going forward.