Skip to main content

Security practices

Tenant isolation​

  • Every record that belongs to a workspace carries the workspace's identifier, and every query filters by the caller's workspace. Records are never read by id alone.
  • The workspace comes from the caller's credential - their session, an API key, an integration - never from a request body or parameter.
  • An automated check fails the build if a new kind of workspace data is added without the workspace marker.

Authentication​

  • No passwords. People sign in through identity providers: built-in ones, or the workspace's own OpenID Connect or SAML provider. A workspace can require its provider to prove multi-factor authentication.
  • Sign-in hardening. OpenID Connect uses PKCE and validates ID tokens against the provider's published keys; the in-progress sign-in is carried in a signed, short-lived, HTTP-only cookie. A completed sign-in is handed to the browser with a single-use code, never a token in a URL.
  • Account linking is conservative. A new provider is linked to an existing account by email only when the provider vouches for the address, and a workspace provider only within domains the workspace has proved it owns by DNS.
  • Sessions. Access tokens last 30 minutes and are refreshed. Roles are looked up on every request, so removing a role takes effect immediately.
  • Participants have no account. A participant credential is good for one session only.
  • Integration launches are signed by the platform's private key, verified against its published keys, single-use, audience-bound and valid for about a minute. A launch is never treated as a sign-in.

The server decides​

Participant identity, permissions, correctness, points, timing and session state are decided on the server. A phone sends answers, never scores or verdicts; an automated check fails the build if any participant request could carry a score or a correctness flag. Answer keys are not sent to phones before the presenter reveals them. Time limits are enforced by the server's clock.

One answer per question, one vote per post and one submission per participant are enforced by unique constraints in the database, not by application checks, so retries and races cannot count twice.

Secrets​

  • API key secrets are stored as hashes; webhook, integration and identity-provider secrets are encrypted at rest. None is ever returned after it is first shown.
  • Integration launch verification is asymmetric: Murmle holds nothing that could mint a platform's users.
  • Webhooks are signed with HMAC-SHA256 over a timestamp and the body, so they cannot be replayed under a fresh timestamp.

Abuse controls​

  • Joining by code, and launches, are rate-limited per network address; pulling outcomes and self-registration are rate-limited per key. Limits are kept in the database, so they hold across every server.
  • Join codes are random, released when a session closes, and reveal only a session's title and a join form.
  • An unknown code, a closed session, an unknown API key and a wrong secret each produce a uniform answer that does not say which it was.

Web protections​

  • Participant-written text is always shown as text, never interpreted as HTML.
  • Only the participant embed view may be framed by other sites, and only by origins the operator allows. Everything else refuses to be framed.
  • Media in content slides must be https://; YouTube videos use the privacy-enhanced domain; nothing autoplays.

Logging and telemetry​

Logs and traces are structured and carry a request id. They never contain survey free text, participant answers, contact fields, tokens or raw participant identifiers, and email addresses are redacted (u***@domain). A rejected credential is logged as rejected, without detail that would help someone forge one.

Audit trail​

Security-relevant actions are recorded in an append-only, per-workspace audit log written in the same transaction as the action, recording what changed but never content.

Data protection​

Retention limits, session and participant erasure, workspace export and deletion, and account deletion are described in Data retention, deletion and export.