These pages describe what the park does today and change when it does. Written by Cedar on 2026-10-07 by reading the server's actual code and settings (not from memory). Items marked to do are gaps found while writing it. COPPA as amended in 2025 (16 CFR 312.8(b)) asks operators to keep a written information security program with a named coordinator, a yearly risk assessment, safeguards, regular testing and a yearly review; this draft is organised that way.
Information security program
1. Who is responsible
- Program coordinator: the operator. Owns this program, approves changes, decides on incidents.
- Day-to-day operation: the server is run from the operator's own account with AI coding assistants (Claude, as "Cedar") working under her direction and her account's permissions. A second assistant ("Oak", on the operator's desktop) builds the 3D models and has no server access beyond dropping files into an upload inbox, by the operator's rule.
- Nobody else has access. There are no employees, contractors or volunteers with access today. [Update this list whenever that changes.]
2. What we protect
| Data | Where | Sensitivity |
|---|---|---|
| Sticker books: random id (UUID), picture name (bean colour + animal), PIN hash, device type, last seen, activity events, stickers, garden, hub visits | SQLite file ~/beansplace-backend/data/beansplace.db | Low: no names, emails, free text or location. The random id is a persistent identifier. |
| Performance reports (device numbers, no identifier) | Same database, perf table | Very low |
| Web server logs (IP addresses) | Hosting log files and monthly archives | Low to medium (IP addresses) |
| Secrets: admin key, mail service keys, mailbox passwords, alert topic, review-queue key | Files in the operator's home folder | High: they open other systems |
| Planned: parent emails, consent records, children's question text, learning-game results | Not built | Medium to high. This program must be updated before any of it is collected. |
3. Safeguards in place today
Collect less
- Kids can play with nothing saved ("Just play"). The backend creates a sticker book only when a grown-up taps Check in (
GET /me?create=0only looks). - No names, emails or free text from children. Event types come from a fixed list; event details are capped at 120 characters and are written by the game, not typed by children.
- Performance reports are sent with
credentials: 'omit', so they never carry the sticker-book cookie.
Identity and the cookie
- The sticker-book cookie
bp_kidholds a random UUID and isHttpOnly; Secure; SameSite=Lax, 400 days. - Optional 4-digit PINs are stored as scrypt hashes with a random 16-byte salt and compared with a timing-safe check. The PIN itself is never stored.
The server
- The backend (Node.js/Express) listens on
127.0.0.1:3010only and is reached through Apache as a reverse proxy. It runs under pm2 and restarts on reboot. - All traffic is HTTPS (Let's Encrypt certificate for beansplace.kids, valid to 2027-01-04, renewed automatically by the hosting control panel).
- Request limits: 120 API requests a minute per IP address; request bodies up to 16 KB; at most 60 events per request.
- The hosting firewall (firewalld) and fail2ban protect the server; the mail server has its own fail2ban jail.
- The play-stats endpoint needs an admin key and returns counts only, nothing that identifies a child.
Secrets
- Every secret lives in a file only the operator's account can read (mode 600): the backend's
.env, the mail service keys, the mailbox passwords, the alert topic and the review-queue key. - Secrets never go into the code repository: the sync script copies code only (never
data/,.envornode_modules/) and refuses to commit anything that looks like a key.
Changes and releases
- Every change is made on a separate dev copy first and goes live only through a deploy script that runs checks, keeps a snapshot of the live site, and rolls back by itself if the live site fails its check. A one-command rollback is always available.
- A guard stops test files, documents and drafts from ever being copied to the live site.
- Standing rules for the AI assistants: never delete backups or people's data without the operator; never touch other servers, DNS, the firewall or money without her.
Watching
- An uptime check every 5 minutes alerts the operator's phone if the site or its API goes down, and again when it's back.
- Each deploy and rollback sends an alert. A morning summary sends counts only.
4. To do (gaps found while writing this)
- Backups: the VPS provider backs up the entire system nightly (2026-10-07). On top of that, a nightly SQLite
.backupof the park's database (consistent even mid-write, integrity-checked) is kept 30 days, readable only by the server account (~/beansplace-tools/backup_db.sh, cron 03:15). [confirm how long the daily server backups are kept] - File permissions: the database files are readable by other accounts on the server (folder 755, files 644). Tighten to 700/600.
- Server logs: the hosting archives each month's access logs and keeps them. Our policy says 14 days. Delete archives after 14 days and ask the hosting provider (or change the control panel setting) about the live logs.
- Sticker books without a PIN: anyone who picks the same picture name on another device can open the newest sticker book with that picture and no PIN. The data is low-sensitivity (activity and stickers, no names), but the check-in should encourage a PIN, or "find my sticker book" should need one.
- Performance reports: they never send the cookie, but the server would store a sticker-book id if one arrived. Make the server ignore the cookie on that route entirely.
- Deletion: there's no deletion tool yet. Build the "Forget this sticker book" button and an operator script that deletes one sticker book and all its rows, and the 12-month clean-up job (first deletions are due 2027-10).
- Written assurances from service providers (312.8(c)): keep a copy of each provider's data-protection terms: hosting [HOSTING PROVIDER]; later Mailjet (parent emails) and Anthropic (question text).
- AI assistant access: write down which commands the assistants may run without asking, and review that list here every year.
5. Yearly risk assessment
Done each [MONTH], and before any new kind of data is collected. The first one:
| Risk | Likelihood | Impact | What we do |
|---|---|---|---|
| Someone breaks into the server | Low | Medium: sticker books, logs, secrets | HTTPS, firewall, fail2ban, backend on localhost only, secrets mode 600, updates via the control panel. To do: tighten database permissions. |
| A sticker book opened by someone who picks the same picture | Medium | Low | PIN option exists. To do: nudge PINs, or require one to find a book. |
| Data lost in a server failure | Low | Medium | Daily server backup + nightly consistent database copy (30 days). |
| An AI assistant makes a harmful change | Low | Medium | Permission prompts for risky commands, dev-first changes, automatic rollback, standing rules. To do: backups (above). |
| A secret leaks into the code repository | Low | High | The sync refuses key-like files and never copies .env; secrets live outside the repository. |
| Logs kept longer than promised | High (today) | Low | To do: 14-day log deletion. |
6. Testing and monitoring
- Automated checks run before every release, including an API test of the backend, and the live site is checked after each release.
- Every quarter: test a backup restore, and confirm the nightly deletion jobs ran (they write to a log).
- Every year: re-read the code paths that store data and update section 2 and the privacy policy if anything changed.
7. If something goes wrong (incident response)
- Contain: take the affected part offline, or roll back the site. Change any secret that may have leaked.
- Assess: what data, whose, how many sticker books, and for how long.
- Tell people: The operator decides who must be told and by when (parents, authorities).
- Fix and record: fix the cause, write down what happened and what changed, and update this program.
8. Yearly review
The coordinator reviews this program at least once a year, and whenever we add a new kind of data, a new service provider, or a new person or assistant with access. Next review: [DATE].