The security model
Plugboard scopes application access by institution and permissions. Production
Compose adds database row-level security through its restricted app_user
connection. Demo and native deployments use an owner connection, so they do not
have that independent database restriction. Connector secrets are encrypted;
passwords and API keys are stored as hashes.
Authentication
Section titled “Authentication”Local accounts use a modern memory-hard password hash.
Staff passwords need at least 14 characters, or 8 where a second factor stands behind them at sign-in, at most 256, and never a common one. The common-password check uses a list carried in the product and makes no network call. See sign-in security.
Sessions use short-lived access tokens with revocable refresh tokens. Refresh tokens are revoked in the same transaction that deactivates an account, whether you do it by hand or through SCIM. A session left idle ends after the time the school sets, 60 minutes for a fresh school; see sessions and signing out.
Every access token names the session it belongs to, and each request checks that session is still live. From 0.55.0, ending a session (Sign out everywhere, a password reset, a replayed refresh token) refuses its access tokens on their next request instead of when they expire.
Each refresh retires the token it used and issues a replacement. A retired token presented again outside the grace window is treated as a possible replay: the account’s live sessions are ended at first detection. Two requests from the same browser can legitimately race, for example when a page reloads mid-refresh, so a request inside a short grace window is handed the same replacement instead. Before 0.16.0 that replacement could be lost, and staff were occasionally signed out of every device without warning; that no longer happens. A reuse outside the window is recorded in the audit log as Refresh token reuse detected, naming the person and how many sessions were ended. From 0.22.1, presenting that same retired token again does not create another alert or revoke later sessions. The token itself is never recorded.
Two-factor authentication is TOTP, set up by each user under their account. Turning it off requires a current code. Each code is accepted once, so a code somebody watched you type cannot be used again. Who must use it is the school’s choice under sign-in security, and a fresh school asks it of administrators. A passkey can stand in for the authenticator where the school says so.
Single sign-on is generic OIDC or SAML 2.0. The OIDC flow uses PKCE and a nonce. See SSO.
Authorisation
Section titled “Authorisation”Role-based, with fine-grained permissions. A role
holding * is a superuser within its own institution, and no permission crosses
into another.
The same checks run on the web console, the REST API, the MCP server and Ask.
Separate permissions allow administrators to distinguish these operations:
ticket.internalseparates the desk’s own notes from working the queue, so you can give somebody the queue without the notes.charge.raiseandcharge.approveseparate proposing a charge from making it real. Waiving sits with approval.
Isolation between institutions
Section titled “Isolation between institutions”Tenant-owned records are accessed through the scoped database wrapper. Some deployment-wide identity lookups deliberately resolve the tenant before a scoped query. See tenant isolation for the deployment differences.
Secrets
Section titled “Secrets”Connector credentials are encrypted before they reach the database and decrypted in memory only for the duration of a call. See secrets and keys.
The outbound fetch guard
Section titled “The outbound fetch guard”Outbound requests to private, loopback and link-local addresses are blocked by default, so a URL typed into a service monitor or a webhook cannot reach an internal host.
A self-hosted install that monitors LAN addresses opts out:
ALLOW_PRIVATE_EGRESS=1The setting allows private and loopback destinations deployment-wide; it never permits link-local or cloud metadata addresses. Leave it off if the deployment serves more than one institution. A managed deployment uses a connector agent instead, which reaches internal systems without Plugboard requesting a private address.
Guarded requests pin approved addresses and validate redirects. The Ollama
connector requires an agent, the operator’s
private-egress setting or the exact configured LOCAL_AI_URL origin for private
access. See Ollama.
Transport
Section titled “Transport”TLS 1.2 is the floor. HSTS is sent, and HTTP redirects to HTTPS.
The desk sends a content security policy for scripts: a fresh nonce for every
page, so a script injected into a page does not run, and no plugins or base-URL
changes. It is sent report-only until a deployment sets
CSP_ENFORCE,
so the browser reports what it would block and blocks nothing; see
operations for the count of reports.
Framing is enforced either way: the desk refuses to be framed except by the
origins in EMBED_ORIGINS.
HSTS comes from the application itself rather than from a reverse proxy in
front of it, so a deployment that terminates TLS somewhere we did not write the
config for still sends it. The default is one year including subdomains, and
HSTS_MAX_AGE changes or removes it.
The API honours X-Forwarded-Proto and X-Forwarded-For from a reverse proxy.
See HTTPS and certificates.
Rate limiting
Section titled “Rate limiting”Applied to authentication endpoints, and to the public API per key.
Audit entries cannot be edited or deleted through the application. Sign-ins,
refused sign-ins, sign-outs and changes to a person’s own password and
authenticator are recorded. Coverage is not complete across all mutations; for
example, API-key lifecycle operations have no entries. Production app_user
also cannot update or delete audit rows; an operator’s database-owner access
has broader powers. See the audit log and API-key
coverage.
Plugboard support reaches a managed deployment with a separate deployment credential, which opens nothing until the school grants support access for a set time. Every support request is recorded with the operator’s name before it reads anything. The grant governs the support channel; it does not change infrastructure administrator access to the server we run.
Supply chain
Section titled “Supply chain”Release container images are signed and include an SPDX software inventory. The image scan fails on fixable high or critical vulnerabilities. Source scans and dependency checks provide additional coverage, not a guarantee that no vulnerabilities remain.
Every download has a checksum published beside it, and the automatic updater refuses an archive that does not match. Verification instructions are in updating.
What is your responsibility
Section titled “What is your responsibility”SECRETS_MASTER_KEY, kept somewhere other than the server.
The certificate, and its renewal.
Who holds which role. The product enforces permissions; it cannot tell you
that four people should not hold device.manage.
Deciding when to update. Nothing updates itself except the on-premises bundle, and you can turn that off.
Restoring a backup once, to confirm your regime works.
Reporting a vulnerability
Section titled “Reporting a vulnerability”Email [email protected] with reproduction details. The repository security policy targets acknowledgement within two business days and initial assessment within five business days. Avoid posting unpatched vulnerability details publicly.