Security
How CyberGuard Academy is built, what is enforced rather than promised, and what we do not claim.
Last updated November 12, 2026
This document is not finished.
The details below have not been filled in yet, so this page is not indexed by search engines and must not be relied on. Each one appears in the text as a visible marker rather than a guess.
- The legal name of the business selling the service.
- A monitored address for privacy requests and legal notices.
- A postal address where a legal notice can be served.
We sell security training, so this page is held to a higher standard than a marketing section. Every measure below is enforced by something specific in the system and covered by an automated check. The last section lists what we do not claim, because a security page without one is asking to be taken on trust.
One company cannot see another's data
This is the property that matters most in a product where several companies' employee records sit in one database, and it is the one most often implemented in a way that looks right and is not.
Isolation is enforced in the database, by row-level security policies, rather than by a filter in application code. The difference is what happens when something is written wrongly: a forgotten where clause in application code returns another customer's rows, while here the database refuses the request regardless of what the application asked for.
Three specific traps are closed and each is covered by a test that fails the build:
Every view runs as the caller. A Postgres view without
security_invokerexecutes as its owner and ignores row-level security entirely — returning every tenant's rows while the policies beneath it look perfect. Every view in this system sets it, and a test asserts it.Access is resolved from live membership records, not from a token. Permissions encoded in a signed token stay valid until that token refreshes, which can be hours. Deactivating an employee here cuts their access on their next request.
The quiz answer key is not granted to signed-in users at all. Revoking a single column would not have been enough: in Postgres, a column-level revoke silently does nothing while table-level access is still held. The tables carry no grant, so the answers cannot be read out of the page or the API.
The audit log is append-only for everyone, including us. Records can be written and read, never edited or deleted.
How we know, rather than believe
The isolation claim above is checked by a suite that signs in as one company's administrator and attempts to reach another company's data — employees, progress, quiz attempts, certificates, simulation results, billing records and reminder history — and fails if anything comes back. It runs against the real database, not a mock. 86 assertions.
Alongside it: a full end-to-end suite that drives the deployed site in a real browser and completes a course, earns a certificate, verifies it, launches a simulation and reads the report; unit tests on the seat and pricing arithmetic; a check that the TypeScript types still match the SQL; and a guard that fails the build if a secret or a service-role import can reach browser code.
Writing those tests has already caught real defects that code review had not, which is the only argument for them worth making.
Phishing simulations are fenced by design
A feature that sends convincing deceptive email on a customer's behalf is one bad decision away from being abusable. The limits are structural, not policy:
Recipients are read from the customer's own employee records. The function that launches a campaign takes no address list, so there is no parameter through which an outside address could be reached.
Message templates are written and owned by us. A customer cannot compose arbitrary content, so the feature cannot be used to send real phishing or marketing mail.
The landing page an employee reaches contains no form of any kind. No password is ever requested, typed or recorded.
There is no tracking pixel and no concept of an 'open'. Only two things are recorded, and both are deliberate human acts: following the link, and reporting the message.
Campaign launches are rate limited per company, in the database.
Payments
Card details are entered on pages hosted by Stripe. No card number, expiry or security code ever reaches our systems, and there is no card field anywhere in the application.
Our payment webhook verifies Stripe's signature against the exact bytes received before anything is acted on, so the endpoint cannot be used to grant a subscription.
Every payment event is recorded by id before it is processed, so the duplicate deliveries a payment processor normally sends cannot apply a change twice.
The quantity billed is computed by the database from the account's own employee records. No endpoint accepts a quantity, and the customer-facing payment portal has seat editing switched off — so a browser cannot influence what it is charged.
The application itself
A Content Security Policy with a per-request nonce and
strict-dynamic, so an injected inline script does not execute. Connections are limited to our own origin, our database provider and Stripe.Authentication, password hashing and session handling are provided by Supabase Auth. We never see or store a password.
Administrative actions are recorded in the append-only audit log with who did them and when.
Sensitive operations — invitations, quiz submission, certificate verification, simulation launches — are rate limited in the database rather than in front-end code.
Secrets live in the hosting environment, never in the repository, and an automated check fails the build if one appears in code or if server-only code is imported into a browser bundle.
What we do not claim
Read this part. It is the most useful section on the page.
No SOC 2, ISO 27001 or HIPAA certification. We have not been audited against any framework. If your procurement process requires an attestation, we do not currently have one, and we will tell you that rather than point at our subprocessors' certifications as though they were ours.
No third-party penetration test. The checks described above are ours. An independent assessment would be worth more, and we have not commissioned one.
No uptime guarantee and no service credits. We do not publish an availability figure because we do not yet measure one properly, and a number we cannot stand behind is worse than no number.
No formal bug bounty, though we will respond to any genuine report and will not pursue anyone who makes one in good faith.
Our Spanish translations are reviewed by a native speaker before employees see them, and until a translation is approved, employees are shown the English with a note saying so. Unreviewed machine translation of security instructions is a safety problem, not a cosmetic one.
Reporting a vulnerability
Write to [to be completed: contact email]. Tell us what you found and how to reproduce it. We will acknowledge it, and we will not take action against anyone who reports a genuine issue in good faith.
This page is a description of how the service works and how we handle information. It is not legal advice, and it does not replace advice from your own counsel about your obligations.