Level 1 completed 25 October 2026, Level 2 on 31 October 2026
This self-assessment covers all 70 Level 1 requirements and the 183 additional Level 2 requirements in the OWASP Application Security Verification Standard 5.0. Each result explains how a requirement is met, or why it is not. The plain-language summary is on the security page.
Current scope: Levels 1 and 2. Level 1 was completed on . The assessment was extended to Level 2 on , replacing the earlier Level 1-only assessment. The Level 1 results include two requirements met by a different route and nine that do not apply. Level 2 is not fully met: 10 requirements are not met, by choice.
This is the developer’s self-assessment, not an independent audit or OWASP certification. The scope, method, results and limitations are published below.
What was assessed, and how
Standard.OWASP ASVS version 5.0.0, Levels 1 and 2: the 70 requirements of Level 1 and the 183 Level 2 adds, taken from the official release file.
Scope. The EquiDiary web application at app.equidiary.com and its API, which the iPhone and Android apps also load, and the static site at equidiary.com, and the off-site backup server that copies their data each night. The platforms underneath — Cloudflare’s, Apple’s, Google’s — are outside it.
Method. For each requirement we read the code that implements it. Where the result can be seen from outside, we also checked the live service, on 25 October 2026 for Level 1 and on 30 and 31 October for Level 2: response headers, cookie attributes, TLS handshakes. Many of the controls are exercised by the automated tests that run before every deploy, against a database built from the real schema. Gaps found along the way were fixed and checked again before this result was recorded; the few we chose not to fix are listed, each with our reasons.
Who. The person who builds EquiDiary. That makes this a self-assessment, not an independent audit; OWASP does not certify or endorse applications.
Result. Level 1: 59 met, 2 met by a different route than the requirement describes (explained below), 9 not applicable. Level 2: 125 met, 2 met by a different route, 46 not applicable, and 10 not met, by choice (explained below).
The requirement names in the tables are our own short summaries; the standard has the exact wording of each.
Everything shown in the app is HTML-escaped where it is inserted, or set as plain text. File-name headers are built in a restricted ASCII form plus RFC 5987; calendar text is escaped per RFC 5545. The Content-Security-Policy forbids inline scripts as a second line.
1.2.2 URLs built safely
Met
No link is rendered from user input. Phone links are filtered to digits and dial characters, research links are built from the DOI over https, and file links are generated by the server.
1.2.3 JavaScript and JSON encoding
Met
JSON is produced only by the framework’s serializer. No script is built from data, and the pages have no inline scripts.
1.2.4 Database injection
Met
Every query binds its values as parameters. The few SQL fragments assembled in code are constants, never input.
1.2.5 OS command injection
Not applicable
The service runs on Cloudflare Workers, which has no operating-system commands to call.
1.3.1 Rich-text sanitization
Not applicable
No HTML or rich-text input is accepted: everything users type is plain text, escaped on output. Research abstracts from outside are reduced to plain text before they are stored.
1.3.2 No dynamic code execution
Met
No eval, new Function or string timers, and the CSP does not allow them.
1.5.1 XML parsers
Not applicable
No XML is parsed anywhere; every integration speaks JSON.
V2 Validation and business logic
Requirement
Result
How it is met
2.1.1 Validation rules documented
Met
A table kept with the code gives the rule for every data item: names (120 characters), notes (4,000), new passwords (8 to 200), real calendar dates, times, prices (0 to 1,000,000,000), ids, lists, and uploads (photos 5 MB, PDFs 10 MB).
2.2.1 Input checked against expectations
Met
Each field is read through a validator that accepts only the expected type, format, range or set — a date that exists, a role from a fixed list, bounded numbers and lengths — and refuses the request otherwise. The database repeats the closed sets as constraints.
2.2.2 Validation on the server
Met
All checks run on the server and in the database. Checks in the app, such as the horse limit or a video’s length, are repeated there.
2.3.1 Steps in order, none skipped
Met
Emailed links work once and are claimed in a single statement. A temporary password allows nothing until it is replaced. A subscription becomes active only when the payment provider says so, never because the browser came back from checkout, and a video counts as ready only when the video service reports it.
V3 Web frontend security
Requirement
Result
How it is met
3.2.1 Content not rendered in the wrong context
Met
Uploaded files are served only for the contexts the app uses them in — as an image or video, on their own, or fetched by the app — and refused when requested as a frame, embed, script or stylesheet (judged by Sec-Fetch-Dest). Photos and videos also carry a sandboxed, deny-everything policy, every response carries X-Content-Type-Options: nosniff, and file types are verified at upload.
3.2.2 Text rendered as text
Met
Text reaches the page through textContent or an HTML-escaping function, everywhere. The one piece of markup a script writes without escaping, the tag that loads the reader’s language, holds nothing but a language code from a fixed list.
3.3.1 Secure cookies
Met
The one cookie is __Host-eq_session: Secure, Path=/, no Domain, HttpOnly, SameSite=Lax. Check: sign in and read the Set-Cookie header.
3.4.1 HSTS for at least a year
Met
Every response carries Strict-Transport-Security: max-age=31536000; includeSubDomains. Check: curl -sI https://app.equidiary.com.
3.4.2 CORS restricted
Met
No Access-Control-Allow-Origin header is ever sent, so no other origin can read a response.
3.5.1 Cross-site request forgery
Met
A request that changes anything must come from the app’s own origin, judged from Sec-Fetch-Site and Origin; sibling subdomains are refused too. Only the two payment webhooks are exempt, being authenticated by signature, and the address browsers send reports of what a page’s security policy refused to, which changes nothing.
3.5.2 No way around the CORS preflight
Met
Writes that carry data as JSON must be sent as application/json, which another site cannot send without a preflight — and none is granted. The rest — file uploads and voice dictation, which are multipart forms, and starting a checkout — could be sent by another site without one, so for those the app does not rely on the preflight: they are refused by the origin check in 3.5.1, and a post from another site carries no session cookie in the first place (SameSite=Lax).
3.5.3 Safe methods stay safe
Met
Every change is a POST, PUT or DELETE. GET requests only read; the one exception notes when a calendar feed was last fetched.
V4 API and web service
Requirement
Result
How it is met
4.1.1 Content-Type with charset
Met
Every text response names UTF-8, and JSON is served as application/json.
4.4.1 WebSockets over TLS
Not applicable
No WebSockets.
V5 File handling
Requirement
Result
How it is met
5.2.1 Upload size limited
Met
12 MB per request; photos 5 MB, PDFs 10 MB, voice clips 4 MB; a cap on files per entry and per horse; videos at most 3 minutes each, enforced by the video service as well.
5.2.2 Extension and content agree
Met
The type must be on an allowlist (JPEG, PNG, WebP, PDF), the file’s first bytes must match it, and the stored extension comes from the verified type. The uploaded file name never names or serves a file; it is kept only as a passport or paper’s title when none is typed.
5.3.1 Uploads never executed
Met
Files are kept in private object storage and only ever streamed back as data.
5.3.2 Safe file paths
Met
Storage keys are generated by the server — area, id, a random part, the extension — and checked against a strict pattern whenever one is read.
V6 Authentication
Requirement
Result
How it is met
6.1.1 Anti-automation documented
Met
The operations documentation sets out the bot check, the attempt counters (per address and network, per network, per account), the edge rate limit, and why a stranger cannot lock someone else’s address.
6.2.1 At least 8 characters
Met
8 minimum and 200 maximum, enforced on the server.
6.2.2 Password can be changed
Met
Account → Change password.
6.2.3 Change needs the current password
Met
The current password is required, and a wrong one counts against the attempt limit.
6.2.4 Common passwords refused
Met
New passwords are checked, case-insensitively, against the 5,000 most common of 8 characters or more, from the UK NCSC’s list.
6.2.5 No composition rules
Met
None.
6.2.6 Password fields masked
Met
Every password field is type=password. The sign-in, registration and administrator set-password fields also have a button to show what was typed.
6.2.7 Paste and password managers allowed
Met
Paste is never blocked, and fields carry the autocomplete hints password managers use.
6.2.8 Password checked exactly as typed
Met
Never trimmed, truncated or case-folded.
6.3.1 Brute-force defences in place
Met
A bot check that fails closed, then an attempt counter claimed in one statement before any hashing: 10 failures per address and network in 15 minutes, 50 per network. Changing the password and deleting the account, which need a session, allow 10 failures per account in the same window. Two-step sign-in codes allow 10 wrong per account in the same window, and after 100 wrong in a row only a recovery code works. An edge rate limit sits in front.
6.3.2 No default accounts
Met
None exists. The placeholder an empty database starts with has a password hash that cannot match anything.
6.4.1 Initial secrets expire
Met
Emailed links are random (128 or 256 bits), expire (a reset after 60 minutes, a confirmation after 2 days) and work once. A password set by an administrator, or generated for a new account, only lets its holder choose their own, and expires after 3 days.
6.4.2 No hints or secret questions
Met
None.
V7 Session management
Requirement
Result
How it is met
7.2.1 Sessions verified by the server
Met
Every API request looks its session up on the server, by a hash of the token, with an expiry check.
7.2.2 Tokens generated per session
Met
A random reference token for each sign-in; no static keys.
7.2.3 Tokens unique, 128 bits or more
Met
256 bits from a cryptographically secure generator.
7.2.4 New token at every sign-in
Met
Signing in issues a new token and ends the one the browser held. Changing the password issues a new token and ends every other; this device’s previous one answers for one more minute, for requests already in flight.
7.4.1 Ended sessions stay ended
Met
Logout deletes the session on the server, and expiry is enforced in the query that reads it.
7.4.2 Sessions end with the account
Met
Deleting an account deletes its sessions and its calendar link. Roles are read afresh on every request.
V8 Authorization
Requirement
Result
How it is met
8.1.1 Access rules documented
Met
Owners, contributors and viewers, per-category sharing, share links and owner-only actions are documented with the code.
8.2.1 Function-level access
Met
Every API route requires a session except four kinds: those that start or recover one (sign-in, registration, the emailed confirmation and reset links, logout, and the check of whether one exists); the two payment webhooks, authenticated by signature; the share-link routes and the virus scan’s verdicts, each authenticated by its token; and the address browsers send reports of what a page’s security policy refused to, which grants nothing and changes nothing. Every administrative route checks the administrator role.
8.2.2 Data-level access
Met
Ownership and sharing are part of every query’s conditions; someone else’s record answers “not found”. A file is checked through the record it belongs to, and an account’s photo by the same rule that decides who may see its name.
8.3.1 Enforced on the server
Met
Every rule is enforced by the server. The app only hides what the server would refuse anyway.
V9 Self-contained tokens
Requirement
Result
How it is met
9.1.1 Signatures checked before use
Met by a different route
Polar’s webhooks are verified (HMAC-SHA256, constant-time compare). Apple’s signed notifications are never trusted at all: see below.
9.1.2 Algorithms on an allowlist
Met
Algorithms are fixed in code. A token’s own algorithm header is never consulted, so “none” cannot be chosen.
9.1.3 Keys from trusted sources
Met
Keys come from server secrets and fixed hosts. No key or key address carried inside a token is followed.
9.2.1 Validity period honoured
Met
Webhook timestamps must be within 5 minutes. Apple’s signing dates order updates, so an old one cannot overwrite a newer one.
V10 OAuth and OIDC
Requirement
Result
How it is met
10.4.1–10.4.5 Authorization server duties
Not applicable
EquiDiary is not an OAuth authorization server. It only uses others: Zoho for mail, and Google, Microsoft and Apple for signing in.
V11 Cryptography
Requirement
Result
How it is met
11.3.1 No ECB or weak padding
Met
Nothing uses ECB or PKCS#1 v1.5 encryption. The service encrypts two things itself: each account’s two-step sign-in key, with AES-256-GCM, and its backups, with age (ChaCha20-Poly1305).
11.3.2 Approved ciphers only
Met
Authenticated encryption only.
11.4.1 Approved hash functions only
Met
SHA-256 for every hash and signature and every MAC but the two-step codes’, and argon2id for passwords — over the old PBKDF2-HMAC-SHA256 hash for an account whose password predates it; no MD5. SHA-1 appears twice: as the index Have I Been Pwned’s breached-password lookup is built on, where it names a range to ask for and protects nothing; and as the HMAC in two-step sign-in codes, the only one authenticator apps make, which NIST still approves as an HMAC.
V12 Secure communication
Requirement
Result
How it is met
12.1.1 Current TLS versions only
Met
TLS 1.2 and 1.3, with 1.3 preferred. Check: a TLS 1.0 or 1.1 handshake with openssl s_client is refused.
12.2.1 TLS everywhere, no fallback
Met
HTTPS on every host and HSTS set. Over plain http, a page is redirected and the API refused. The platform’s default workers.dev address is switched off.
12.2.2 Publicly trusted certificates
Met
Certificates from Google Trust Services.
V13 Configuration
Requirement
Result
How it is met
13.4.1 No source-control metadata exposed
Met
Only the public folder is deployed: /.git/ and the configuration files answer 404.
V14 Data protection
Requirement
Result
How it is met
14.2.1 No secrets in URLs
Met by a different route
The session, passwords and bot-check tokens travel in headers and bodies, and emailed links carry their token in the URL fragment, which browsers never send. Share links and calendar feeds carry a token in the address by necessity: see below.
14.3.1 Device cleared at the end of a session
Met
Logout deletes the offline journal, the queue of unsent changes and the cached photos before it contacts the server; offline, the session itself is ended at the next launch. A launch that finds no session clears them too, except changes typed on the device and never sent: those stay, and nothing else of the journal does, until the same account signs in again, and another account signing in deletes them.
V15 Secure coding and architecture
Requirement
Result
How it is met
15.1.1 Remediation time frames documented
Met
A high or critical advisory in code that ships is fixed within 7 days, a moderate one within 30, and dependencies are updated monthly.
15.2.1 Nothing past its deadline
Met
Five packages are dependencies, twelve with what they pull in. On 28 October 2026, npm audit reported no known vulnerability in any of them.
15.3.1 Only the fields needed are returned
Met
Responses name the fields they return, account data passes through one allowlist, and prices are masked for anyone not allowed to see them — even when they edit an entry.
Level 2
The requirements Level 2 adds to Level 1, assessed on 31 October 2026. Where one is not met, the row says so, and Not met, by choice gives our reasons.
V1 Encoding and sanitization
Requirement
Result
How it is met, or why not
1.1.1 Input decoded once
Met
Each piece of input is decoded once, by the part of the service that expects that encoding, before anything checks it: the address of a request, read in the one form every check and the routing share, the name of a stored file, and the text of research records from outside.
1.1.2 Encoding where it is used
Met
Text is stored as it was typed and escaped at the moment it is used: for the page when it is shown, for an email or the research page when it is built, for a calendar line when it is written, and for a file name when the download header is made.
1.2.6 LDAP injection
Not applicable
The service queries no LDAP directory.
1.2.7 XPath injection
Not applicable
No XML is parsed and no XPath is run; every integration answers JSON.
1.2.8 LaTeX injection
Not applicable
The service runs no LaTeX.
1.2.9 Regular expressions from data
Not applicable
No regular expression is ever built from data, so there is nothing to escape.
1.3.3 Cleaning for risky places
Met
Each place where text could do harm has its own cleaning and length limit: names are kept to one line and 120 characters before they reach an email subject or a calendar, download names lose control characters and slashes, calendar text is escaped and folded as the standard asks, and text from outside services is reduced to plain text and checked before it is stored.
1.3.4 SVG images
Met
SVG files are never accepted. Uploads must be JPEG, PNG, WebP or PDF, and their first bytes must match the type they claim, so an SVG renamed as a PNG is refused.
1.3.5 Markdown, CSS and templates
Not applicable
No Markdown, stylesheet, BBCode or template language is accepted from anyone: everything is plain text, and colours come from a fixed palette.
1.3.6 Server-side request forgery
Met
The service never fetches an address a user gives it. Every outgoing request goes to a fixed list of hosts over https, a redirect is refused unless that service is known to send one (and then each step is checked), and the identifiers placed in those requests are checked against a strict pattern first.
1.3.7 Template injection
Met
No template is built from input. Emails and screens are fixed texts with named gaps that values fill, and a check fails any translation that adds a gap of its own.
1.3.8 JNDI injection
Not applicable
The service has no Java, and so no JNDI.
1.3.9 Memcache injection
Not applicable
No memcache is used. The one cache is keyed by the site’s own address and a language from a fixed list.
1.3.10 Format strings
Met
The code has no format strings. Values are placed into texts in a way that takes them literally, so a name containing special characters cannot change the text around it.
1.3.11 Email injection
Met
Mail is sent through the mail provider’s web API, not SMTP. Each message goes to one checked address, its subject is always one line, and every value in its body is escaped.
1.4.1 Memory-safe code
Not applicable
Everything is TypeScript and JavaScript, which manage memory themselves; none of the five libraries that ship contains native code.
1.4.2 Integer overflow
Met
Every number is checked against a range before use: ids, prices (up to a billion, so no total overflows), durations, repeat intervals, link lifetimes, video trims and page sizes. Requests for part of a file are matched to a strict pattern and kept within the file.
1.4.3 Freed memory
Not applicable
The runtime frees memory itself; there are no pointers to leave dangling.
1.5.2 Safe deserialization
Met
The only format read is JSON, which creates no objects of a chosen type. Each field is read by name, a request is never merged into another object, and lookups use only an object’s own keys.
V2 Validation and business logic
Requirement
Result
How it is met, or why not
2.1.2 Rules across fields documented
Met
A table kept with the code lists every rule that ties two values together, such as a care plan that repeats or falls on one date but not both, and where each is enforced.
2.1.3 Limits documented
Met
The limits are written down, those per account (horses, attachments, video, storage, changes a day, emails, dictation) and those for the whole service (sign-ups per network, calls to the video service, the security log’s ceilings).
2.2.3 Related values checked together
Met
The server checks that values make sense together: an entry’s horse, person and type, a person limited to some types, training-only fields, a trim within its video, a restricted share that names at least one type, and categories at most two levels deep.
2.3.2 Limits enforced as documented
Met
Each documented limit is enforced on the server, and the ones that cap what an account adds are checked inside the very statement that adds the row, so parallel requests cannot slip past together. Tests race them.
2.3.3 All or nothing
Met
Each operation that writes several things writes them in one database transaction: an account and its first records, a password change, a share and the types it covers, a person and their types, a share link and its scope, and an account’s deletion. A file stored before its record is removed again if the record is refused.
2.3.4 No double-booking
Met
Everything limited is claimed in a single statement: emailed links and invitations work once, daily allowances and per-minute rations cannot be overspent by parallel requests, one export runs at a time, and one App Store subscription belongs to one account.
2.4.1 Anti-automation
Met
A bot check guards sign-in, registration and password resets, one network opens at most ten accounts a day, and wrong passwords are limited. Each account may make 2,000 changes a day, and calls to the video service, the payment provider and Apple are rationed per account and for everyone; storage, video, dictation, emails, messages and exports have their own limits, and so does the security log.
V3 Web frontend security
Requirement
Result
How it is met, or why not
3.3.2 SameSite set for the cookie’s purpose
Met
The session cookie is SameSite=Lax: a browser sends it when someone follows a link into the app, and never with a request another site makes in the background or a form it posts. That suits it, because nothing changes on a plain page load and every change must come from the app’s own address (3.5.1). It is the only cookie the app sets.
3.3.3 __Host- cookie prefix
Met
The session cookie is named __Host-eq_session, so the browser keeps it to the one host that set it and no other subdomain can set or replace it. No cookie is meant to be shared. The old unprefixed name is still accepted for sessions signed in before the prefix arrived, and swapped for the prefixed one on first use; those sessions end within 90 days of their sign-in.
3.3.4 Session cookie out of scripts’ reach
Met
The session cookie is HttpOnly, so no script on the page can read it, and the session token reaches the browser only in the Set-Cookie header: no answer of the API carries it. The list of signed-in devices names each session by a short code the token cannot be worked back from.
3.4.3 Content Security Policy
Met
Every page carries one policy that names where scripts may come from (the app itself and Cloudflare’s bot check), with object-src 'none' and base-uri 'none', no inline script and no eval. Stored photos and videos carry a stricter, sandboxed policy of their own. Check: curl -sI https://app.equidiary.com.
3.4.4 nosniff on every response
Met
Every response the server sends carries X-Content-Type-Options: nosniff: pages, the API’s answers and refusals, scripts, stored files, redirects and the answers to plain http. Only the responses Cloudflare’s network makes itself, such as its rate-limit and error pages, lack it.
3.4.5 Referrer policy
Met
Every page carries Referrer-Policy: strict-origin-when-cross-origin, so another site learns at most which site a visitor came from, never the page: the address of a share link stays private. Links mailed for a reset or a confirmation carry their token after a #, which browsers never send anywhere. A PDF opened on its own carries no policy of ours and gets the browser’s default, which in current browsers is the same one.
3.4.6 Framing refused
Met
Every response says frame-ancestors 'none' in its Content Security Policy, so no other site can show any part of the app inside a frame. Nothing of ours is meant to be embedded.
3.5.4 Separate applications on separate hosts
Met
The app is at app.equidiary.com, the website at equidiary.com and the test version at its own address behind a sign-in; videos play from the video service’s own addresses. The session cookie belongs to the app’s host alone and the website refuses to serve the app, so the website never holds a session. The same server answers both addresses, so a share link also opens on the website’s address and the website’s pages on the app’s; all are our own pages under the same policy.
3.5.5 postMessage origins checked
Met
The app’s one listener for messages is its offline worker’s, which keeps the reader’s language for use without a connection: it ignores a message from any other site, and any message that is not the address of a language file. None of the app’s pages listens for messages from other windows, and the bot check’s own script handles the messages from its own frame.
3.7.1 Only supported client technology
Met
Plain JavaScript and standard browser features (a service worker, IndexedDB, audio recording). No plugin, Flash, Java applet or Silverlight anywhere, and the page policy forbids plugins outright.
3.7.2 Off-site redirects allowlisted
Met
The server redirects only within its own addresses, and the app takes no redirect address from a link. The one place the app sends someone to another site by itself is subscribing or managing a subscription: the address comes from the payment provider’s own service, and the server passes it on only when it is a secure address on the provider’s domain.
V4 API and web service
Requirement
Result
How it is met, or why not
4.1.2 Only pages redirect http to https
Met by a different route
Over plain http a page is sent on to its https address, while the API and the calendar feed are refused with no redirect, so a program that sends a password or a token unencrypted fails where its developer sees it. Browsers never get that far: they are told to use https only. The one exception is a calendar link handed out for one-tap subscription before the app stopped offering it: those links are marked, their owners are asked to make a new one, and until they do a calendar’s plain-http request for one is redirected (details).
4.1.3 Proxy headers cannot be forged
Met
The server trusts only the fields Cloudflare’s network writes over whatever a visitor sends: the connecting IP address, its country and the request’s id. It reads no X-Forwarded-For, X-Real-IP or user header.
4.2.1 Request smuggling
Not met, by choice
No server of ours sits behind the network: Cloudflare’s edge and its runtime read every request before the app sees it, and the app passes no raw request on to another server. That parsing is Cloudflare’s engineering, which nothing in our code can test.
4.3.1–4.3.2 GraphQL
Not applicable
No GraphQL, and no client sends a query: every filter is a fixed, checked parameter, a search is cut to 200 characters and matched as plain text, and one answer holds at most 1,000 entries.
4.4.2–4.4.4 WebSockets
Not applicable
No WebSockets.
V5 File handling
Requirement
Result
How it is met, or why not
5.1.1 Upload rules documented
Met
The documentation lists every upload: photos (JPEG, PNG or WebP, 5 MB), PDFs (10 MB), videos (3 minutes a clip, 200 MB a file, handled by the video service) and voice dictation (4 MB, never stored). No upload accepts an archive, so nothing is unpacked. It also says how a stored file is served so it cannot run, how it is checked for viruses, and what happens to one that is caught.
5.2.3 Archives checked before unpacking
Not applicable
No upload accepts a compressed archive, and the server unpacks nothing. The one zip, the account export, is one the server builds.
5.4.1 Download file names
Met
A stored file is named by the server, never by the uploader, and every file is served under a name the server made. A file downloaded under a readable name gets that name cleaned, and keeps the extension of its checked type. The uploaded file’s own name is kept only as a wallet document’s title when none is typed, and never names a file.
5.4.2 File names encoded
Met
Download names are written as RFC 6266 asks: control characters and slashes removed, a plain-ASCII version in quotes and the full name encoded as UTF-8 beside it. Other file names are letters, digits, dots, dashes and underscores only. No email carries an attachment.
5.4.3 Virus scanning
Met
Every photo and document is checked with ClamAV, on current signatures, the night after it is uploaded. A file it matches is withheld from then on, wherever it would be served: to its owner, through share links and in the account export, and a device that had shown it checks again within a day. Until that first check a file can be opened, for a day or a little more. Videos are only ever served as the video service’s own re-encoding, never as uploaded.
V6 Authentication
Requirement
Result
How it is met, or why not
6.1.2 Words to avoid in passwords documented
Met
The security inventory lists the words a guesser tries first: the app’s name, horse words in the app’s nine languages, and the account’s own address and names.
6.1.3 Every way in documented
Met
The security inventory lists every way into the service, what each one grants and what guards it: sign-in and its two-step code, registration, the emailed links, what an administrator can do to an account, share and calendar links, the payment providers’ notifications and the virus scan’s report.
6.2.9 Long passwords allowed
Met
A new password can be up to 200 characters. Signing in sets no limit, and no form cuts short what is typed.
6.2.10 No forced password changes
Met
A password never expires. Only one an administrator set has a deadline: its holder must replace it with their own within 3 days.
6.2.11 Context words refused
Met
A new password that is little more than those words with digits and symbols added is refused, at registration, reset, change and when an administrator sets one.
6.2.12 Breached passwords refused
Met
New passwords are checked against Have I Been Pwned’s list of breached passwords, which receives only the first five characters of a hash. If that check cannot be reached, the password is allowed and the common-password list still applies.
6.3.3 A second factor for everyone
Met by a different route
Two-step sign-in, a code from an authenticator app or a security key, is offered to every account and required of administrators. Why it is not required of everyone, and what stands in for it: see below.
6.3.4 No undocumented ways in
Met
Every request needs a session unless its route is on a short public list, each one documented. With two-step sign-in on, every way to a session needs the code; a reset or an administrator setting a password creates no session.
6.4.3 Reset cannot skip two-step
Met
The reset link is random (256 bits), works once and expires after 60 minutes. Using it signs out every device, disconnects any Google, Microsoft or Apple sign-in and signs nobody in, and two-step sign-in stays on, so the next sign-in still asks for the code.
6.4.4 A lost second factor
Met
Recovery codes cover a lost phone. With those lost too, an administrator turns two-step sign-in off only on a request confirmed from the account’s own address, after typing their own password and code, and the owner is emailed. Signing in still takes the password, which is all that setting it up took.
6.5.1 Codes work once
Met
An authenticator code is accepted once, and no code for the same or an earlier 30 seconds is taken after it. A security key signs a random challenge that works once, and a key that counts its uses must count on. Each recovery code works once.
6.5.2 Recovery codes stored safely
Met
Each recovery code carries 120 random bits, above the 112 at which a plain hash is enough, and only its SHA-256 hash is stored.
6.5.3 Codes from a secure generator
Met
The authenticator key (160 bits) and the recovery codes come from a cryptographically secure generator.
6.5.4 Codes long enough
Met
Recovery codes carry 120 bits. No code is sent by email or text.
6.5.5 Codes short-lived
Not met, by choice
Level 2 asks an authenticator code to be accepted for 30 seconds at most. The service also accepts the 30 seconds before and after, up to 90 in all, for a phone whose clock is a little out. Each code still works once, and wrong codes are limited.
6.6.1–6.6.3 Codes by phone, text or email
Not applicable
There are none. The second factor is an authenticator app or a security key, with recovery codes.
6.8.1 One provider cannot speak for another
Met
A sign-in with Google, Microsoft or Apple finds the account by the provider and that provider’s own identifier for the person together, never by the email address. An address that already has an account is refused, never joined to it: its owner connects the provider from Account, after typing their password again.
6.8.2 Signed sign-in assertions
Met
Every sign-in token from Google, Microsoft or Apple has its signature checked against that company’s published keys before anything in it is read, with the one algorithm they use, whatever the token claims; an unsigned token is refused.
6.8.3 SAML assertions used once
Not applicable
No SAML.
6.8.4 Strength claimed by the provider
Met
The service never relies on the provider for how strong a sign-in was: two-step sign-in, when it is on, is asked after Google, Microsoft or Apple as after the password. A provider is used to confirm it’s you only for an account that has no password; an account with a password is always asked for it. Where the provider says when its user last signed in, a sign-in older than 10 minutes is refused; where it does not say, its answer is taken to mean only that the browser could sign in to the connected account at that moment.
V7 Session management
Requirement
Result
How it is met, or why not
7.1.1 Session lifetimes documented
Met
A session lasts 90 days from sign-in and ends after 30 days unused. The documentation says why that is longer than NIST’s guidance (the app is used offline at the barn, and often only every few weeks) and what stands in for shorter sessions.
7.1.2 Parallel sessions documented
Met
Unlimited, one per device. Every one is listed under Account → Signed-in devices, and none is ended to make room for another.
7.1.3 Federated sessions documented
Not applicable
No single sign-on or identity federation.
7.3.1 Inactivity timeout
Met
A session unused for 30 days stops working, checked on every request. The app keeps no session alive on a timer, except to ask whether a video it is showing has finished converting.
7.3.2 Absolute session lifetime
Met
Every session ends 90 days after its sign-in, however much it is used. Nothing extends it.
7.4.3 Sign out other devices after a change
Met
Changing or resetting the password, or an administrator setting one, signs out every other device by itself. Turning two-step sign-in on or off, or making new recovery codes, offers a button to sign out the others, and Account → Signed-in devices offers it at any time.
7.4.4 Log out always within reach
Met
Account is in the top bar of every signed-in view, with Log out in it. The screen that makes someone replace a temporary password has its own Log out.
7.4.5 Administrators can end sessions
Met
An administrator can sign one account out of every device, or every account but their own current session, after typing their password and code again.
7.5.1 Sensitive changes need a fresh sign-in
Met
No account can change its own email address. Changing the password, deleting the account, every change to two-step sign-in and handing a horse over to another account ask for the password, and with two-step on a code as well. An administrator changing another account’s address types their own password and code again first.
7.5.2 Sessions can be seen and ended
Met
Account → Signed-in devices lists every session with its browser, system, country and last use. Signing one out, or all the others, asks for the password again, and the code with two-step on.
7.6.1 Federated re-authentication
Not applicable
No outside identity provider.
7.6.2 Sessions only by an explicit action
Met
A session is created only by signing in, typing the two-step code or registering, each a form the person submits. Emailed links create none, and visiting the app signed out sets no cookie.
V8 Authorization
Requirement
Result
How it is met, or why not
8.1.2 Field-level rules documented
Met
A table kept with the code says, for every kind of data, who sees it: the owner, a contributor, a viewer, a share-link guest, any other account and an administrator. It includes what a share’s own settings hide (prices, the wallet, the details of a horse’s trips, the feeding plan), the owner’s own entry types, which everyone a horse is shared with sees, and when another account’s name and photo may be shown. Who may change what is written beside it: the feeding plan, the wallet and the care schedules are the owner’s alone, a writer kept from prices cannot set one, and a task changes only by whoever raised it or the horse’s owner.
8.2.3 Field-level access enforced
Met
Each of those rules is part of the database query that reads or writes the data. A price, a wallet document, or another account’s name and photo never leaves the server for someone the owner has not allowed, even in the reply to their own edit. The share-link page returns exactly what it names: no prices, authors, wallet or tasks, and the feeding plan only when the link carries it.
8.4.1 Accounts kept apart
Met
Every account is its own tenant. Each query carries the owner or sharing condition, and another account’s records answer “not found”. Deleting an account keeps what it logged on other owners’ horses and hands its own entry types to those owners. Once a helper’s access to a horse has ended, they can no longer rename the people or entry types of their own that label its owner’s entries.
V9 Self-contained tokens
Requirement
Result
How it is met, or why not
9.2.2 Token type and purpose checked
Met
A payment webhook changes a subscription only after its signature checks out, and only for subscription events. An App Store notification is used only to learn which subscription to look up, and of Apple’s answer only the transaction is read, for our own products. Each of the service’s own tokens is kept apart by what it is for, so a token made for one purpose works for no other.
9.2.3 Token audience checked
Met
Each payment environment has its own webhook secret, so a delivery signed for another endpoint fails. Apple’s answer must name our app and one of our products.
9.2.4 Audience in the tokens we issue
Met
Each key we sign with serves one audience. The App Store key signs only requests to Apple’s subscription service, naming our app, for 30 minutes. The video key signs only playback tokens, each for one video, for 12 hours.
V10 OAuth and OIDC
Requirement
Result
How it is met, or why not
10.1.1 Tokens only where needed
Met
Signing in with Google, Microsoft or Apple runs on the server: the service’s secrets and the sign-in token never reach the browser, and no access token from them is kept. One long-lived token is kept: Apple’s, for each account connected to Apple, because Apple expects the service to cancel the connection on Apple’s side when that account is deleted. It is stored encrypted, is used for nothing but that cancellation, and is worth nothing without the service’s own Apple key, which is not in the database. The one-time code only passes through the browser on its way to the server (in the iPhone app, from Apple’s own sign-in sheet; for Google and Microsoft there, it stays in the phone’s own browser sheet and never reaches the app), and is worth nothing without those secrets. The service is also a server-side client of Zoho’s mail, kept as the way back for the mail Cloudflare’s email service now sends. Its long-lived credentials are server secrets sent only to Zoho’s sign-in address, and the short-lived access token goes only to Zoho’s mail address, for one batch of mail. None of them is stored, logged or sent to a browser.
10.1.2 Flow bound to the session
Met
Each trip to Google, Microsoft or Apple carries its own random state, nonce and code verifier, and is tied to the browser that set out by a cookie: the provider’s answer is taken only when that browser brings it back, and only that browser can read how it ended, once, within 10 minutes. In the iPhone app, Google and Microsoft sign in through the phone’s own browser sheet, which is a second browser, so the same tie is held in three parts: the sheet starts only when the app itself opens it, never from a link sent to someone; the provider’s answer is taken only from that sheet; and how it ended can be read only by the app that opened the sheet, which is handed a one-time ticket through an address that only EquiDiary’s app can receive. Connecting a provider or confirming it’s you also comes back only to the session that asked. Zoho’s grant was made once, by hand, and is only renewed.
10.2.1 Forgery protection on the code flow
Met
PKCE for Google and Microsoft, and a random state for all three; Apple, which does not take PKCE, by its state and nonce.
10.2.2 Mix-up attacks
Met
Each provider has its own return address, and a code is only traded with the provider its trip was started for; each sign-in token’s issuer must be that provider’s.
10.3.1–10.3.4 Resource server duties
Not applicable
The API accepts no OAuth access token: only its own session cookie, its own link tokens and the payment webhooks’ credentials.
10.4.6–10.4.11 Authorization server duties
Not applicable
EquiDiary is not an OAuth authorization server. As a client, each token request carries its client secret and asks only for what it uses: from Zoho the two permissions sending mail needs, from the sign-in providers the person’s identifier, email address and name.
10.5.1 Sign-in tokens not replayed
Met
Each trip sends a random nonce that the sign-in token must carry, and is spent by its outcome.
10.5.2 The person identified for good
Met
An account is found by the provider’s identifier for the person, which the provider never hands to anyone else, and never by the email address or the name.
10.5.3 The provider is who it says
Met
Each token’s issuer must be the provider’s own, fixed in the code; nothing about a provider is looked up at run time.
10.5.4 The token is for this service
Met
A token must name this service as its audience, must not have expired, and must have been issued within the hour.
10.5.5 Back-channel logout
Not applicable
Signing out at a provider does not sign anyone out here: EquiDiary keeps its own sessions.
10.6.1–10.6.2 OpenID Provider duties
Not applicable
EquiDiary is not an OpenID Provider.
10.7.1–10.7.3 Consent management
Not applicable
There is no authorization server, so there is no consent to collect, show or revoke.
V11 Cryptography
Requirement
Result
How it is met, or why not
11.1.1 Key management policy
Not met, by choice
Keys are managed through the lifecycle NIST SP 800-57 describes (generation, distribution, storage, use, rotation, destruction), and each key’s lifetime and what to do if it leaks are written down. The test environment holds a copy of four production credentials (the App Store key, the mail grant, and the video service’s token and signing key), one holder more than the requirement allows: see below.
11.1.2 Cryptographic inventory
Met
A table kept with the code lists every algorithm, its settings, its key and where it is used, the site’s certificates included. A change that adds a key must update it, and the assessment is redone at least once a year.
11.2.1 Established implementations
Met
Cryptography is done by the platform’s built-in Web Crypto (BoringSSL), by the noble-hashes library for argon2id, and by age for the backups. Nothing cryptographic is written by hand.
11.2.2 Algorithms and keys can be changed
Met
Each stored password names its algorithm and settings and is upgraded at the next sign-in, as the move from PBKDF2 to argon2id was. The key that protects two-step keys can be replaced: the old one keeps opening them while each is sealed again under the new one. Every other key has written rotation steps.
11.2.3 At least 128-bit strength
Not met, by choice
Everything we choose reaches 128 bits: argon2id, HMAC-SHA256, AES-256-GCM, P-256, X25519, 160-bit two-step keys and 128- or 256-bit tokens. The video service’s RSA signing key, and the RSA certificate our hosting provider offers a client that asks for one, are sized by those providers at about 112 bits: see below.
11.3.3 Encrypted data protected from tampering
Met
Both things the service encrypts use authenticated encryption: each two-step key with AES-256-GCM, bound to its account, and the backups with age.
11.4.2 Password hashing
Met
argon2id with 19 MiB of memory, 2 passes and a random salt, which is OWASP’s recommended minimum. An older hash is upgraded at the next sign-in, or wrapped in argon2id by a nightly job.
11.4.3 Hashes strong enough for integrity
Met
Every signature and MAC uses SHA-256. SHA-1 appears only as the HMAC in two-step codes, which needs no collision resistance, and as the breached-password service’s index, which protects nothing.
11.4.4 Keys derived from passwords
Not applicable
No key is derived from a password: passwords are only ever checked, and backups are encrypted to a public key.
11.5.1 Random values
Met
Every secret comes from the platform’s cryptographically secure generator: 256 bits for sessions and emailed links, 128 bits for share, calendar and invite links, 160 bits for two-step keys. Recovery codes, at 120 bits, follow the authentication chapter’s own rule for them.
11.6.1 Approved signature algorithms
Met
Signatures use ECDSA P-256 and RSA PKCS#1 v1.5, both with SHA-256, on keys Apple and Cloudflare generate; the sign-in providers’ tokens are checked as RSA PKCS#1 v1.5 with SHA-256 only. The backup key comes from age’s own key generator.
V12 Secure communication
Requirement
Result
How it is met, or why not
12.1.2 Strong cipher suites only
Not met, by choice
Under TLS 1.2, our hosting provider’s default still accepts a few suites without forward secrecy and a few CBC suites with SHA-1, beside the strong ones that current browsers negotiate first. Restricting them is a paid add-on we have not bought.
12.1.3 Client certificates checked
Not applicable
No connection is authenticated with a client certificate.
12.3.1 Encryption on every connection
Met
Every connection into and out of the service uses HTTPS, and the backup server is administered over SSH. A plain-http request is redirected to https if it is for a page and refused if it is for the API.
12.3.2 Certificates checked
Met
Every outgoing connection checks the server’s certificate, and nothing switches that check off, on the service or on the backup server.
12.3.3 Encryption between our own parts
Met
The one network link between our own parts, the backup server reporting to the service, is HTTPS. The database, the file store and the speech service are reached inside the platform, not over a network.
12.3.4 Trusted certificates between our own parts
Met
Those links use publicly trusted certificates. We run no certificate authority of our own and no self-signed certificate.
V13 Configuration
Requirement
Result
How it is met, or why not
13.1.1 Connections documented
Met
Every outside service the server talks to is listed, with why, how the connection is authenticated and what it sends, and so is everything that connects in. An automated check fails when the list and the code differ. No address a user types is ever fetched.
13.2.1 No long-lived keys between components
Not met, by choice
The payment, mail, bot-check, video and summary services issue only long-lived keys, and the backup server reaches its storage and the service with long-lived keys too. Each is held by one party, limited to what its use needs, and replaced every year; the backup server’s keys to our hosting provider work only from that server’s addresses.
13.2.2 Least privilege between components
Met
Each key can do only what its use needs: the payment key creates checkouts and billing-portal sessions, reads prices and ends the subscription of an account being deleted, and the backup’s key to the file store only reads. The backup runs as a user of its own rather than as the server’s administrator, can write only its own folders, and is handed its keys in memory for each run.
13.2.3 No default credentials
Met
No service uses a default password or key. Every secret is generated at random or issued by its provider.
13.2.4 Allowlist of outside services
Met
The server can reach only a fixed list of outside hosts, over HTTPS, and follows no redirect except where a service is built to send one, checking each step against the list. An automated check refuses any outgoing request that bypasses it.
13.2.5 Server held to that allowlist
Met
The list is enforced in the one function every outgoing request goes through, the summary service’s library included. Files and data are read only from the platform’s own storage.
13.3.1 Secrets in a secrets manager
Not met, by choice
The service’s secrets are in the hosting platform’s secret store, which cannot be read back, and none is in the code. The backup server keeps its keys in a file only its administrator can read, rather than in a vault; the backup gets a copy in memory for each run, and its keys to our hosting provider work only from that server.
13.3.2 Least privilege for secrets
Not met, by choice
The test environment uses the same video, mail and App Store accounts as production, so its keys reach production’s. It deploys only what the owner pushes to it and answers only behind an access check.
13.4.2 No debug mode in production
Met
There is no debug mode, debug route or source map in production. An unexpected error answers “Internal error”, and the details go only to our own logs.
13.4.3 No directory listings
Met
No folder is ever listed: a folder’s address shows the not-found page, or on the app’s address, the app.
13.4.4 No TRACE
Met
The service answers no TRACE request, and the hosting platform refuses the method before it arrives.
13.4.5 No internal documentation or monitoring exposed
Met
No API documentation is published, and the internal documents are not deployed. Every administrative and monitoring route needs an administrator who has passed two-step sign-in.
V14 Data protection
Requirement
Result
How it is met, or why not
14.1.1 Data classified
Met
Everything the service holds is classified as secret, confidential, internal or public, down to two-step sign-in keys, video playback tokens, the addresses a horse is shared with and the IP address passed to the bot check. The classes follow the GDPR, the UK GDPR and Mexico’s data protection law.
14.1.2 Protection requirements per class
Met
Each class has written requirements for transport, use, logging, encryption at rest, integrity and how long it is kept.
14.2.2 Not cached on servers
Met
API answers are marked never to be stored, and files as for the reader’s own browser only, so no shared cache keeps them. The only pages the server caches are the public research digest’s.
14.2.3 No trackers
Met
There is no analytics service or tracker, and the only outside script is the bot check. Personal data goes only to the companies the privacy notice names.
14.2.4 Protection applied as documented
Met
Each class’s requirements are what the code does: secrets stored only as hashes or encrypted, addresses kept in logs only in scrambled form, accounts logged by number, nothing logged while an address carrying a link’s token is served, and each store emptied on its stated schedule.
14.3.2 Not cached in browsers
Not met, by choice
API answers, the data download, the calendar feed, wallet documents and a share link’s files are never stored by the browser. Photos and PDF attachments are kept in its cache, and shown for a day without asking again, until signing out clears that cache.
14.3.3 Nothing sensitive in browser storage
Not met, by choice
The app keeps the journal, the changes still to send, and the photos and PDFs it has shown on the device, because it is used at the barn without signal. All of it is wiped at sign-out and when the app opens without a session; wallet documents and videos are never stored there.
V15 Secure coding and architecture
Requirement
Result
How it is met, or why not
15.1.2 Software inventory
Met
Every library that ships is listed, five direct and twelve in all, each pinned by version and checksum and fetched from the public npm registry. The build tools, the one outside script and the backup server’s packages are listed with their sources.
15.1.3 Costly functions documented
Met
Each expensive function is listed with what bounds it: the data download, password hashing, dictation, uploads, sign-up, the video service’s shared call budget and the research digest among them. Long work is streamed or split, so no answer outlasts its caller.
15.2.2 Costly functions bounded
Met
Each bound is enforced: bot checks, attempt limits, daily caps per account, one data download at a time, size limits, and a per-account and an overall ration of calls to the video service, which also pauses all but deletions when that service says the shared budget is spent.
15.2.3 Nothing extra in production
Met
Tests, scripts and internal documents are neither deployed nor served, the platform’s preview addresses are off, and the test environment’s shortcuts do nothing in production.
15.3.2 Redirects not followed
Met
Outgoing requests follow no redirect, except a video download, where the video service sends one by design; each step is held to the list of allowed hosts.
15.3.3 No mass assignment
Met
Every field a request may set is read by name and checked; nothing else in a request reaches the database.
15.3.4 Real client IP
Met
The client’s address is taken only from the header the hosting platform sets and overwrites, and one a client sends is ignored. Limits per network count an IPv6 address by its /64.
15.3.5 Types checked, comparisons strict
Met
The server is type-checked in strict mode, every input is checked for its type before use, and comparisons are strict.
15.3.6 No prototype pollution
Met
Keys a client chooses are looked up only among an object’s own properties, collections use maps, and request data is never merged into an object.
15.3.7 No parameter pollution
Met
Request bodies are JSON only, each value is read from one named place, and a repeated parameter counts once.
V16 Security logging and error handling
Requirement
Result
How it is met, or why not
16.1.1 Logging inventory
Met
The security documentation lists every place a log is kept: the security log, the account log, the attempt and request counters, the platform’s log of the server’s output, the alert mail, the daily check-in, the encrypted backup, the backup server’s own logs, and the hosting provider’s record of the requests its rate limit blocks. For each it gives the format, what is in it, who can read it and how long it is kept, and it defines every security event recorded.
16.2.1 Entries say when, who, where and what
Met
Each security record notes the time in UTC, the account it concerns and the administrator acting, the route and the records it reached for, the network (in scrambled form) and the country, the outcome, and the request’s ID, which finds the same request in the platform’s log.
16.2.2 Times in UTC
Met
Every log stamps its time in UTC, the backup server’s included, and the admin screen shows each time in UTC and says so.
16.2.3 Logs go only where documented
Met
The server writes logs only to the stores the inventory lists, and every outbound connection is checked against a fixed list of hosts that the build holds to the documentation. No log goes to an analytics service.
16.2.4 Logs can be read and joined
Met
The platform keeps each line as a JSON event, and security records are printed with named fields. Every record of a request carries the request’s ID, which joins the database log, the platform’s log and the error alerts.
16.2.5 Sensitive data kept out of logs
Met
Passwords, tokens and keys are never logged. Email addresses and networks are kept only in scrambled form, and accounts by number. Error text from other services has addresses masked, and nothing is printed while a share or calendar link is served, because the platform files every line under the address it was printed for. Query strings are redacted.
16.3.1 Every sign-in event logged
Met
Registrations, sign-ins (naming the second factor used), a correct password waiting for its code, failed sign-ins, lockouts, logouts, password changes and resets, the password typed again before a sensitive action, two-step changes and wrong codes, recovery codes, and another server’s credential refused are all recorded.
16.3.2 Failed authorization logged
Met
Every refusal for lack of a right is recorded, and so is every “not found” answered to a signed-in user, which is how another account’s records answer, with what was reached for. So are links that match nothing, a share-link visitor asking for a file the link does not cover, and an App Store subscription claimed by a second account.
16.3.3 Documented events and bypass attempts logged
Met
The documented list of events is the list the code writes, and the build fails if the two differ. Input refused, failed bot checks, lockouts, limits reached, oversized requests, cross-site writes and dead links are all recorded, within an hourly cap per source that stops a flood from filling the log.
16.3.4 Errors and failed controls logged
Met
Unexpected errors are logged with the route, the account and the request’s ID, and mailed to operations once an hour per kind. A service that fails, its connection refused or its answer wrong, is logged in its own words, and a security record the database would not take is still printed.
16.4.1 No log injection
Met
Records are written as bound parameters and hold codes, numbers and scrambled values rather than free text. The platform keeps each line as one JSON event, so a value cannot start a second entry. Alert mail and the admin screen escape what they show.
16.4.2 Logs protected
Met
Only administrators who have passed two-step sign-in can read the security log, and no answer contains a scrambled value. The database refuses any change to a record, and any deletion before its 90 days. The platform’s log is readable only by the account’s one member, who signs in with a second factor, and the server can add to it but not change it.
16.4.3 Logs copied to a separate system
Met
Every security record is printed to the platform’s log as it is written, a store the application can add to but not change or empty, so a takeover of the application could not erase that copy. The whole log also leaves every day, encrypted, for a backup server at another provider, and alerts go to a separate mailbox and a dead-man’s switch. Records made while a share or calendar link is served are the exception: they reach the backup the next day.
16.5.1 Generic error messages
Met
An unexpected error answers “Internal error”, and a database conflict a general sentence. Every message is written in the code, and another service’s own words go only to the log. No stack trace, query, key or token is returned.
16.5.2 Outside services failing safely
Met
A provider outage switches off its feature, never a control. The bot check fails closed; the breached-password check gives up after 3 seconds and the common-password list still applies; a video whose playback link cannot be issued is shown without it; the research digest falls back to a second AI model and its pages to the published copy; the video service is left alone for five minutes after it refuses; alert mail is throttled.
16.5.3 Failing closed
Met
Every check refuses when it cannot decide: the bot check, the attempt limits (claimed before any password is checked), sessions, webhook signatures and the virus scan’s token. Limits are tested inside the database write that would pass them, and a change in several steps is written all at once or not at all.
V17 WebRTC
Requirement
Result
How it is met, or why not
17.1.1 TURN server
Not applicable
No WebRTC, so no TURN service. Videos are uploaded over HTTPS and played in the video service’s player; dictation records a clip and uploads it.
17.2.1–17.2.4 WebRTC media server
Not applicable
No media server, DTLS or SRTP.
17.3.1–17.3.2 WebRTC signaling
Not applicable
No signaling server, and no WebSockets.
Met by a different route
Links that carry a token (14.2.1)
Three kinds of address hold a secret because the address is the product. A share link is something an owner sends to a buyer or a clinic. A calendar feed is what Google Calendar, Apple’s Calendar or Outlook subscribe to, and they keep only an address. A video’s playback link is signed by the video service, which puts the signature in the path.
What protects them: share and calendar tokens are 128 random bits, stored only as a hash, read-only, and can be revoked or replaced at any moment; a share link can also expire and be limited to some categories or a start date. A video’s link is valid for at most 12 hours and for one video. The service prints nothing to its logs while serving these addresses, query strings are redacted from the logs, and the referrer policy keeps a page’s address out of the requests it makes to other sites.
One weakness remains, and it belongs to Level 2 (4.1.2): a calendar app handed a webcal:// address may make its request over plain http, and calendar apps ignore HSTS. The app now hands out only the https address and says where each calendar takes it pasted, rather than offering one-tap subscription. A link made before that was offered as webcal://, and a calendar subscribed with it keeps asking for that address until the link is replaced. The app marks those links and asks their owners to make a new one; until they do, a plain-http request for one is redirected to https, while a plain-http request for any other calendar link is refused, like the rest of the API.
Apple’s signed payloads (9.1.1)
Two of Apple’s signed objects reach the server, and the contents of neither are taken from anyone untrusted. A notification about a subscription is read only to learn which subscription to look up; nothing in it is stored. The state is then fetched from Apple’s own API, so a forged notification can at most cause a fresh read of the truth. The transaction inside that API answer arrives over TLS from Apple’s fixed address, in reply to a request we signed, and is held to our app’s bundle id and product list.
Checking Apple’s certificate chain as well would meet the requirement to the letter. We decided against it: it would add a certificate parser to the code that ships, and change nothing about what is trusted.
Two-step sign-in, offered to everyone and required of administrators (6.3.3)
Level 2 asks for a second sign-in factor on every account, and allows an application to relax that with a documented reason and the controls that stand in for it. Two-step sign-in, a code from an authenticator app or a security key with recovery codes beside it, is offered to every account and required of every administrator, whose tools do not open without it.
We did not make it compulsory for everyone. Many owners sign in rarely, often at the barn, and a lost phone without its recovery codes would lock them out of their horses’ records. What stands in for it: wrong passwords are limited per address and network, a new password is checked against the most common and the breached ones, a session ends after 30 days unused, and changing the password or deleting the account asks for the code whenever two-step is on.
Not met, by choice
Ten Level 2 requirements are not met. Each is a decision we recorded, with what limits it:
The journal on the device (14.3.3). Level 2 asks for no sensitive data in browser storage. The app keeps the journal, the changes still to send, and the photos and PDFs it has shown on the device, because it is used at the barn without signal. All of it is wiped at sign-out and when the app opens without a session; passports, papers and videos are never stored there.
Photos and PDFs in the browser’s cache (14.3.2). The browser may show a journal’s photos and PDF attachments from its own cache for a day without asking again, and keeps them until it clears its cache or you sign out, which tells it to. Passports and papers, a share link’s files and everything else the service answers are never cached.
Older TLS 1.2 cipher suites (12.1.2). Our hosting provider’s default still accepts, under TLS 1.2, a few suites without forward secrecy and a few CBC suites with SHA-1, beside the strong ones. TLS 1.0 and 1.1 are refused, and a weaker suite is used only by a client that offers nothing better. Restricting them is a paid add-on we have not bought.
Two RSA keys below 128 bits (11.2.3). The video service signs playback links with an RSA key whose size it chooses, and our hosting provider offers an RSA 2048 certificate to the rare client that cannot use its modern one; both give about 112 bits of strength where Level 2 asks for 128. A playback link works for 12 hours at most.
HTTP parsing is the platform’s (4.2.1). There is no server of ours behind the edge: the hosting platform parses every request, and nothing passes one on to a second parser that could read it differently. That rests on the platform, which we cannot test.
Some keys held twice (11.1.1). The test environment was given production’s keys for the App Store, the mail service and the video service rather than its own, so each has one holder more than the requirement allows.
Staging reaches production’s services (13.3.2). For the same reason, the test environment’s keys can do what production’s can at those three services. It deploys only what the owner, or a tool the owner has given access, pushes to it, and answers only behind an access check.
Long-lived keys between services (13.2.1). Several of the services the app uses issue only long-lived keys, and the backup server reports a virus match with one too. Each key has one purpose and one holder, is limited to what it is used for, and is replaced on a schedule.
The backup server’s keys in a file (13.3.1). The off-site backup keeps its keys in a file only the system administrator can read, rather than in a dedicated vault. The backup runs as its own user, which never reads that file; the keys work only from that server’s address; and everything the backup keeps is encrypted to a key that never touches the server.
Codes valid a little longer (6.5.5). A two-step code is also accepted in the 30 seconds before and after its own, as the standard behind these codes recommends for phones whose clocks drift. Each code works once.