docs: SECURITY.md sagde noget, dagen havde gjort forkert

"Ingen ratebegrænsning ud over login og nulstilling" var sandt i morges. Det er
det ikke længere: andenfaktoren, beaconet og indgående post har fået deres eget
loft i løbet af dagen — det sidste tilføjede jeg selv og skrev ikke det her.

En SECURITY.md, der siger noget forkert, er værre end de fleste fejl. Den er
netop den fil, nogen læser for at afgøre, om de tør køre det her, og hvert eneste
punkt i den er en påstand, de ikke selv kan efterprøve uden at læse koden.

Samme sted mangler den største åbne ting: notekroppe, opgavetitler, beskrivelser
og kommentarer er ikke krypteret. Forseglingen dækker legitimation — postkassers
kodeord, OAuth-tokens, andenfaktorens hemmelighed, AI-nøgler — og intet andet. En
note med et kodeord i har det i klartekst i hver sikkerhedskopi, man kan hente.
Det stod i overleveringen og på dokumentationssiden for noter, og det hørte
hjemme her.

README's "ikke færdigt"-liste havde Gmail med. Den kører i produktion, inklusive
flagningen og dubletbehandlingen, der tog en dag at få rigtig.

Co-Authored-By: Claude Opus 5 <[email protected]>
This commit is contained in:
Kristian Wind
2026-08-20 22:16:52 +02:00
co-authored by Claude Opus 5
parent 264f48ae11
commit 23bd784b07
2 changed files with 27 additions and 6 deletions
+9 -3
View File
@@ -104,15 +104,21 @@ Apple Reminders, forward mail into it, and add it to Claude as a connector.
| AI layer (Anthropic/OpenAI/Google/local) | Built, degrades when unconfigured |
| Gmail: OAuth2 with PKCE, polling, one-way to tasks | Done, tested |
| Google Calendar: read-only, laid over the task calendar | Built, never run against Google |
| Calendar subscriptions: ICS and webcal, several per person | Done, tested |
| Times on tasks, set by dragging in the week grid | Done, tested |
| Notes: archive, grouped list, images, source view | Done, tested |
| Update notice for administrators | Done, tested |
| Personal API tokens | Done, tested |
| OpenAPI 3.1 spec, checked against the router | Done, tested |
| End-to-end smoke tests (Playwright) | Done |
| Documentation, landing page, licence | Done |
**Not finished:** Web Push, the AI providers, and the Gmail and Google Calendar
API calls are implemented against their specifications and have only ever been
exercised there, not against live services. Everything else has an interface: opening a task shows
**Not finished:** Web Push, the AI providers and the Google Calendar API calls are
implemented against their specifications and have only ever been exercised there,
not against live services. Gmail is no longer in that list — it runs in production,
including the flagging and de-duplication that took a day to get right. Calendar
subscriptions over ICS and webcal are proven; the Google half of the calendar is
not. Everything else has an interface: opening a task shows
its description, sub-tasks, comments, files and reminders, and the settings pages
cover the account, notifications, integrations, AI, API tokens, import, export and
templates.
+18 -3
View File
@@ -39,8 +39,23 @@ reply within a week.
- **The CSP** is computed at startup by hashing the inline scripts in the built
page, rather than allowing `unsafe-inline`.
- **Note bodies, task titles, descriptions and comments are not encrypted.** The
sealing above covers credentials — mailbox passwords, OAuth tokens, TOTP secrets,
AI keys — and nothing else. A note holding a password is holding it in plain text
in every backup that can be downloaded. Encrypting note bodies is on the list and
is larger than it sounds: the full-text index and the folded search column are
both derived from the body, so both would have to go and search would have to be
rewritten to decrypt in application code.
## What it does not do
verdande has no rate limiting beyond the login and password-reset endpoints, and
no audit log of reads. It is built to be run behind something — a Cloudflare
Tunnel, a reverse proxy, a VPN — by somebody who knows what they exposed.
Rate limiting covers the endpoints a stranger can reach: signing in, resetting a
password, the second factor, the installation beacon, and inbound mail. **Nothing
else is limited** — an authenticated caller can ask as often as they like, and a
signed-in account that wanted to be a nuisance can be one.
There is no audit log of reads. Who *changed* something is recorded per project;
who looked at it is not.
It is built to be run behind something — a Cloudflare Tunnel, a reverse proxy, a
VPN — by somebody who knows what they exposed.