YaqeenDocs
Back to Yaqeen

Help

Security

The security controls that are built into Yaqeen today. No certification is claimed.

This page lists the controls the product has today. Yaqeen holds no security certification, such as SOC 2 or ISO 27001, and this page does not claim an independent security review. The public endpoint GET /api/v1/trust/controls/ lists the same controls with a pointer to the evidence for each. Last reviewed: 2026-11-19.

Access#

  • Roles. Four roles (reviewer, admin, merchant and system) limit what each account can read and change. See Roles.
  • Sessions. Session cookies are HttpOnly and SameSite=Lax, and Secure in production. Every write needs a CSRF token. Production redirects to HTTPS and sends an HSTS header.
  • Passwords. A password needs at least 12 characters, cannot be a common password, cannot be all digits and cannot be too similar to the username. Staff choose their own passwords through an invitation or a reset link, so an admin never sees one.
  • Sign-in limits. Sign-in is limited per address, per username with address, and per username across all addresses. An IPv6 address counts by its /64 network.
  • Sign in with Google. When it is switched on, people can sign in with a Google account. The API checks Google's signed token, matches by Google's account id and never by email, and never links an email that already has a password account. A new person gets an account that stays switched off until an admin approves it. Google sees that a visitor opened the sign-in page. The audit log records the sign-in with the user id and no email address, because the log can never be erased.
  • Sign-up. Anyone can ask for an account, but it stays switched off until an admin approves it. Sign-up is limited per client and pauses when too many accounts are waiting. A username cannot contain @, so nobody can take someone's email address as a username before its owner signs in with Google. The email given at sign-up is not confirmed.
  • Site headers. The site's web server sends X-Frame-Options: DENY and a frame-ancestors 'none' policy, so no other site can frame the sign-in page. It also sends nosniff, Referrer-Policy: same-origin, a permissions policy that turns off camera, microphone, location and payment, and a content security policy that allows scripts, styles, fonts and images from the site's own address only. The sign-in and sign-up pages also allow Google's sign-in script. Fonts are served from the site itself, so a visitor's browser does not contact Google to show a page. The docs site sends the same set of headers.
  • Claude and Codex access. Sign-in uses OAuth with PKCE S256 only. Access tokens last 60 minutes. Refresh tokens last 30 days, work once, and an old one used again revokes every token of that person for that app. Tokens, codes and API keys are stored as hashes. A token works at the MCP address only, merchant accounts are refused, and the tools are twelve reads. An admin can switch MCP off or turn on redaction, and each call is audited as mcp.call without its arguments. See Connect Claude and Codex.
  • Second-person approval. A large or risky change waits for a different person. See Second-person approval.

Keys and secrets#

  • Scoped API keys. A key has scopes (read, decide, ingest, scim, mcp), an expiry of up to 365 days and its own rate limit. No key can reach accounts, keys, webhooks, privacy requests, usage or sandboxes. Rotation keeps the old key for 24 hours.
  • Hashed secrets. Yaqeen stores only a SHA-256 hash of an API key, an invitation token, a reset token, and the sign-in codes and tokens of Claude and Codex. Each is shown once.
  • Webhook secrets. A webhook signing secret is shown once but stored as text, because Yaqeen needs it to sign each call. Calls carry an HMAC-SHA256 signature and a timestamp. Addresses must use https and must not point inside a private network.
  • Hashed identifiers. National IDs, phones, emails and bank accounts arrive as keyed hashes. The API refuses a raw value.

Records and data#

  • Audit chain. Every change writes a hash-chained, append-only line. This is tamper-evident, not tamper-proof. See How the audit chain works.
  • Tenancy. Each organization sees only its own data, enforced in the default database query. See Organizations.
  • Privacy requests. Access and erasure requests have a due date, and legal holds block erasure. See Privacy and trust.
  • Appeal documents. Files sit in a private bucket. Signed upload forms and download links expire in 5 minutes, and each opening is audited. Uploads are limited to PDF, JPEG and PNG, 10 MB each, five per appeal.
  • Error tracking. The built-in tracker records the error type and the URL pattern, never request data.

Operations#

  • Throttles. Limits apply to sign-in, decisions, data loads, password changes, the assistant and the MCP address, and to each API key per minute and each plan per day.
  • Backups. A daily database backup is encrypted before it is uploaded, and the private key is kept offline. Each dump is restored into a scratch database and checked before old copies are deleted.
  • Retention. Assistant chats, finished webhook deliveries and error records age out by the daily privacy sweep. See Retention.
  • Assistant gateway. A language model, when one is switched on, sees only redacted text. See AI and your data.
  • Alerts and checks. Failed jobs, a broken audit chain and a failed backup raise an alert. A daily job verifies the audit chain.

Security and fairness pack#

A security and fairness pack is available on request. It covers what is not built yet, the results of our fairness checks on the score, and our plans for each. Ask your Yaqeen contact for it.

Report a problem#

To report a security problem, contact your Yaqeen contact directly. Do not include keys, tokens or passwords in the message.

Up Down to moveEnter to openEsc to close