Trust / Security
How it is built, and where it falls short.
This describes the development environment, which is the only one deployed today. There is no production environment yet, so nothing here describes one. Each statement is what the configuration says; it is not an audit. See also the latest assessment status on the trust page.
What is in place
- Secrets live in Azure Key Vault. Pods read them with workload identity, so no static key sits in the cluster. Access is by role, and deleted secrets can be recovered for a period.
- Our pipeline signs in to Azure with short-lived tokens, not stored passwords, and a secret scan runs on every push.
- Images are private and verified. The cluster pulls from a private registry with its own identity. Images are pinned by digest and signed, and a cluster policy checks the signature before an app namespace will run them.
- The database has no public address. It is a managed PostgreSQL server reachable only from inside our network. Backups are kept for a short window (a week) and are not copied to another region.
- A restore has been rehearsed, with receipts kept in the repository. The first two attempts failed and were fixed; the third restored the database to a point in time and verified it.
- One public entry point. The service is reached through a single gateway over HTTPS, behind Cloudflare. Nodes have no public address.
- The operating system of the nodes is scanned and findings are tracked with owners and review dates.
Known gaps
- The Kubernetes control plane accepts connections only from two authorised addresses (restricted on 26 October 2026; before that it was open to the internet). The list is kept by hand and has to be updated when an administrator's address changes.
- The cluster is not yet integrated with an identity provider (Microsoft Entra). Administration uses the cluster's local credentials, which are not tied to named people.
- Approvals are tied to a tenant, not yet to a named person.
- The tamper-evident event log covers the file and in-memory stores; database-backed stores are not chained yet.
- The node operating system has open critical vulnerabilities awaiting a vendor image: see the assessment status. For those with no fix we have written individual assessments from the vulnerability descriptions and what the cluster runs; we have not yet confirmed them against the node's loaded kernel modules.
- Backups are not geo-redundant and the retention window is short.
We list gaps here because a page that lists only strengths would tell you nothing. If you find one we have not listed, tell us.