Security

Tenant isolation at the database, and again in every query

Every table holding your vendors, certificates, documents, templates, reviews and audit history carries an organization id and has PostgreSQL row-level security enabled and forced. The connection is scoped to your organization before any query runs, so a query that forgets its filter returns nothing rather than somebody else's rows — it fails closed. The application connects as a database role that row-level security binds, and refuses to start if it is ever given one that could bypass it. Every query also names your organization explicitly, so isolation never rests on one mechanism alone. Cross-organization checks run in our test suite against a database role configured exactly like production.

A few tables are deliberately outside that policy, and we would rather say so than round the claim up. Each one has to be readable before we know which organization a request belongs to, or exists to answer a question that spans all of them. Support and contact messages can be sent by people who are not signed in, so those rows may have no organization to scope to. Vendor upload links must be resolvable from the token alone — that lookup is by unguessable token, and every action it permits is then re-scoped to the one vendor it names. Membership rows are how we learn which organization you belong to in the first place, and are read by your user id. The email suppression listis deliberately global: a hard bounce or a spam complaint must stop mail everywhere, not only where it happened. The background job queue is claimed by the worker before it knows whose job it is, and the job binds its organization before it reads any of your data. None of these tables holds an insurance document, and none is reachable from a signed-in session without an organization filter in the query.

The AI never decides anything

Extraction is strictly one-way: the model transcribes a certificate to structured data and has no tools, no network, and no authority. Ordinary code then validates that data and scores it against your requirements. Nothing downstream reads free text from a document, so instructions hidden inside an uploaded PDF — a real attack against document-processing systems — cannot influence a compliance decision. We have a test that asserts exactly this.

Your documents are not used to train AI models.

Uploads are treated as hostile

File type is determined by inspecting the bytes, not by trusting a filename or a declared content type. Storage keys are generated server-side and cannot be steered by a filename. Documents live in private storage and are served only through short-lived signed links — there is no public URL for a certificate.

Authentication and access

Sign-in is by emailed magic link, or Google SSO where enabled. A password is optional and additive: you can set one, and nothing stops working if you never do. When you do, we store only a scrypt hash — never the password — and a failed attempt gets one identical refusal whether the address exists, has no password, or has the wrong one, so the refusal cannot be used to enumerate accounts. Setting, changing, resetting or removing a password invalidates every existing session and emails the account; a first set or a reset requires fresh proof (a recent sign-in, or a code emailed to the address on file). Link tokens are stored only as hashes and are single-use. Sessions are HMAC-signed with a sliding window and an absolute expiry. Access is deny-by-default at a single middleware choke point, and every write is checked against your verified role on the server; the browser is never trusted to say what role it has.

An audit log you cannot quietly edit

Consequential actions — a review approval, a compliance decision, a document deletion, a waiver granted — are written to an append-only audit log. Update, delete and truncate are revoked on that table at the database level for the role the application connects as, so no entry can be altered or removed while your account exists, and an entry survives the deletion of the thing it describes: delete a certificate and the record that you deleted it stays.

Append-only is not the same as permanent, so we will be exact about the one thing that ends it. The log is scoped to your organization, and if the account owner deletes the account, it is destroyed along with everything else the organization owns — we do not keep a history of your company after you have told us to delete it. What deliberately outlives the account is a single record that the deletion happened and that any subscription was cancelled; it is kept outside your organization's data precisely so the cascade cannot take it, holds none of your vendors' information, and stores the owner's email address only as a one-way hash. The privacy page lists exactly what that record contains.

Payments

Card details are handled entirely by Stripe on Stripe-hosted pages. We never see, store, or transmit a card number. Subscription state is written only by Stripe's signature-verified webhook — never by anything a browser sends us.

Infrastructure

Traffic is TLS-only with HSTS, a content security policy that admits a script only by a nonce issued for that one response, and strict referrer and content-type headers. Data is encrypted at rest by our database and object-storage providers. Rate limits apply to sign-in, the public vendor portal, and inbound email, and a per-organization daily processing ceiling caps what an abusive sender can cost.

What we are still working on

We would rather tell you than let you assume. Currently in progress: optional two-factor authentication, formal third-party penetration testing, and a data processing agreement you can sign. Our subprocessors are already published — every outside service that touches your data is named, with what it does, in section 5 of the privacy policy, and we post a change there before a new one starts processing your data. If you need any of these before you can adopt CrewCert, write to us and we will tell you honestly where it stands.

Reporting a vulnerability

Email [email protected]. Please include enough detail to reproduce the issue, and give us a reasonable chance to fix it before publishing. We will acknowledge your report, keep you updated, and credit you if you want the credit. We will not pursue legal action against good-faith research that respects customer data and avoids service degradation. Machine-readable contact details are at /.well-known/security.txt.

Privacy

What we collect and why is on the privacy page; data questions go to [email protected].