Security and assurance

Trust and security: how we secure the running platform and check each control automatically and verify and sign and roll back releases and what we test and how you report a vulnerability to us.

On this page
  1. How we secure the platform
  2. Controls in force
  3. Verification before release
  4. Recovery
  5. Tested attacks
  6. Security baseline
  7. Data and outside services
  8. Reporting a vulnerability

How we secure the platform

Four properties that keep a fault containedA failure in the web process must not become control of the host, the database, the release process or another customer's data.
  • A small exposed surfaceWhat can be reached from outside is kept small.
  • Authority at the moment of useA grant is checked when the operation runs. The default answer is refusal.
  • Untrusted content stays inertContent never acquires authority by being stored or read.
  • A record that can be verifiedEvery consequential action leaves an entry that can be checked afterwards.

We build the platform so that a fault stays contained and a failure in the web process never becomes control of the host or the database or the release process or the data of another customer. Four properties carry that: we keep the exposed surface small and we check authority at the moment of use and untrusted content never acquires authority and every consequential action leaves a record that can be verified afterwards.

We enforce each control below in the running system instead of stating it in a policy document and automated tests exercise each one and they must pass before we build a release.

Controls in force

ControlWhat it enforcesHow we verify it
Authority at executionAn operation runs only while a grant for that operation on that resource is current and the default answer is refusal.Automated tests of allowed and refused callers and expired and revoked grants and callers acting on resources that are not theirs.
Sign in and sessionsPasswords and authenticator codes and passkeys. We throttle repeated failures and a session can be revoked and it ends when it expires.Automated tests of sign in and passkey replay and concurrency and throttling and session expiry.
Browser protectionsOur content policy forbids inline and third party scripts and refuses framing and blocks other sources by default. It permits scripts and images and styles only from our own origin and the contact page also permits map tiles from OpenStreetMap. File type protection and a transport security header and origin checks on requests that change state and protected session cookies complete the set.Automated tests of the headers on every response including errors and of the origin checks.
Database privilegesThe application connects as a role that does not own the tables so it cannot rewrite the evidence chain or change a ledger balance directly.Automated tests that run as that role and expect the refusals.
EvidenceConsequential changes write their entry in the same transaction so the change and its record commit together or not at all. Run expiry and approval resolution commit first and record their entry immediately afterwards and the run reports whether that entry was written. Entries are append only and linked by fingerprint so an alteration breaks the chain.Automated tests of the chain and of its verification.
Memory safetyThe code forbids unsafe Rust across the whole project and a build check fails if the setting is removed.The build check that runs before every release.
DependenciesEvery third party component must declare a licence and we generate a software bill of materials from the dependency lock and we declare the outside services that the code can reach.Build checks that fail when a component declares no licence or when the notices or the bill of materials no longer match the lock file or when a scan finds code that makes web requests or opens network connections or resolves names or sends mail without a declared service.
Release and rollbackWe install a release as a numbered release only after a verification receipt exists for its commit. If the new release fails its readiness check the installer points back at the previous release automatically and we apply database migrations before the switch and never reverse them.The installer enforces this for every release on our own host.

Verification before release

From source to a running releaseThe binary that runs cannot come from code the gate did not see.
  1. Clean source treeThe verification gate runs against it.
  2. ReceiptNames the commit and its tree. A release is built only from a tree whose receipt matches.
  3. Signed statementCarries the digest of the software bill of materials that travels beside it.
  4. InstallerRefuses a statement that is unsigned, altered, expired or older than the one in place.
  5. Readiness checkA release that fails it is replaced by the previous one automatically.

The verification gate runs against a clean source tree and leaves a receipt that names the commit and its tree. We build a release only from a tree whose receipt matches and the installer refuses a release without one so the binary that runs cannot come from code that the gate did not see.

We sign release statements with a key held on our host and each one carries the digest of the software bill of materials that travels beside it so the holder of a release can check the binary and its dependency inventory against what we signed. The installer refuses a statement that is unsigned or altered or expired or older than the one in place and it replaces a release that fails its readiness check with the previous one automatically.

Recovery

The recovery drill of 30 October 2026Measured on our own host from our newest encrypted backup.
  • 7.4 secondsTo restore the production data set into a fresh database cluster.
  • Less than a dayThe age of the backup that was restored.
  • 24 hoursThe bound on data loss, set by the nightly backup interval.

Our backups are encrypted and taken nightly and checked instead of assumed. Automated tests restore a backup into an empty database and check the checksum and the schema and every audit chain and the authenticator key. Our own host decrypts its newest encrypted backup and restores it into a fresh database cluster and checks that the stored rows are present. The drill on 30 October 2026 restored the production data set in 7.4 seconds from a backup less than a day old and the nightly interval bounds data loss at 24 hours.

Tested attacks

We keep a register of the attacks that we design against with the postcondition that must hold after each one and they include a stolen session and a compromised web handler and a malicious uploaded document and an injection attempt and a forged customer identifier and a replayed financial operation and a stolen signing key and a compromised host and high volume traffic and more. Sixteen of the eighteen cases carry executed evidence in the register including the hostile document case in which a document that carries a script and an instruction is stored as inert text and reads only inside the account that saved it.

Every release also answers its own acceptance probes from outside and they cover the health and readiness routes and the schema head and every page that the sitemap lists and the refusal of an account page to an anonymous caller and an unknown address that does not echo the address asked for.

Security baseline

We work to a written security baseline that we adopted on 25 October 2026 and it maps to 140 owned requirements and the 345 controls of OWASP ASVS 5.0.0 and we gate every release against the register that records them. When an independent assessment is completed the claim it supports will name the baseline and the scope and the date and the exclusions in the form "assessed against this baseline for this scope on this date".

Data and outside services

Our privacy notice states what personal data we collect and why and for how long. We list the services that handle data outside the application in our sub processor register and we set out the services that the application itself can reach on the deployment and ownership page.

Reporting a vulnerability

Follow our disclosure process and it names the contact address and what to include and the safe harbour that applies to good faith research.

Published
24 October 2026
Revision
14

Last reviewed 13 November 2026.