Security
How the Canvas Grading Assistant protects instructors’ accounts and students’ records, and how to report a security problem.
Sign-in and access
- University accounts only. Sign-in uses the university’s Google (NetID) login, with the standard OpenID Connect authorization-code flow and PKCE. Only university-managed
wisc.eduaccounts are accepted. Personal Google accounts that use a wisc.edu address are refused, and unverified email addresses are rejected. - Sessions last up to 12 hours. They’re stored server-side, and the browser only holds a random token in a secure, HTTP-only, same-site cookie.
- Your own Canvas permissions. Each user connects their own Canvas token, and the Service acts only as that user. Tokens are sent only to approved Canvas addresses (https://canvas.wisc.edu), are never sent to the browser, and are never written to logs.
- Isolation between users. Every record is tied to the signed-in user, and every database query is limited to that user. Automated tests check that one user can’t read or change another’s runs, drafts, text or credit.
- Admin access to usage and billing reports is limited to named administrators.
Protecting data
- In transit: HTTPS everywhere, with HSTS. Certificates are managed by Google.
- At rest: Canvas tokens and extracted submission text are encrypted by the application (AES-256-GCM) before they reach the database, in addition to Google Cloud’s default storage encryption.
- Secrets (encryption key, database password, sign-in and payment keys) live in Google Secret Manager. Only the app’s own service account can read each one.
- Database: Cloud SQL (PostgreSQL) in the U.S., with no authorized external networks. The app connects through Google’s authenticated connector. Daily backups are kept for 7 days.
- Least privilege: the app runs as a dedicated service account that can only call the AI model, connect to its database, read its own secrets and write logs. It has no other project permissions, and no long-lived service-account keys exist.
- Data minimization: only submission content and rubrics go to the AI model, never names or IDs. Photo metadata (location, camera) is removed before transcription. Original files aren’t stored. Extracted text is deleted after 30 days.
- Logs record requests and errors but not submission text, tokens or student names.
Safeguards on grading
- Nothing posts without approval. Every grade sent to Canvas was approved by a person. Posting asks for confirmation, and editing an approved draft clears its approval.
- Flagged items (unreadable, blank, scanned, wrong format, low-confidence, missing work) never get a score automatically.
- Every deduction needs a comment before it can be approved.
- Instructor-of-record sign-off is required per course before drafting.
- Audit trail: every draft, edit, approval and post is logged with the user, time and model.
- Submissions are treated as data, not instructions: the AI is told to ignore instructions written inside student work, and its output is validated against the rubric (scores are capped at each criterion’s maximum).
Payments
Card payments happen on Stripe’s hosted page. We never see or store card numbers. Payment notifications from Stripe are verified by cryptographic signature, and each payment is credited once, only to the account that made it. The app’s Stripe key is restricted to creating and reading its own checkout sessions.
Reporting a security problem
If you believe you’ve found a vulnerability, or think your account or data has been exposed, email support@grader.courses with “Security” in the subject. Please don’t access other users’ data or disrupt the Service while testing. We’ll acknowledge reports within 3 business days.
If we learn of unauthorized access to student records or other personal information, we’ll investigate promptly, notify affected users and the University as required by law and University policy, and take steps to contain it.
Contact
Laura Montoya · support@grader.courses
Last updated November 16, 2026.