Security
What protects your workspace, your callers and your payment details, and the parts you control.
Last updated 13 November 2026
On this page
The short version
Callaroo separates each business's data in the database, encrypts traffic and stored data, keeps card numbers with Stripe, and checks carrier and payment webhooks before acting on them. This page also names the gaps we know about, including unfinished database enforcement, an admin sign-in check that looks at enrollment rather than a fresh check, and a text check that misses some scam instructions. We do not hold a published third-party security certification.
Read the sections below for detail. Notes headed 'What this means for you' say what to do with the facts in each section.
How the service is put together
Callaroo is a web dashboard plus a backend that handles telephony and voice sessions. The dashboard reads business data through an authenticated, user-scoped connection. The backend holds the carrier and voice-model credentials and is the only part that talks to Telnyx and the voice model provider. Provider keys are not shipped to the browser.
That separation matters: a bug in a page cannot hand out a carrier key, and the browser never receives a service credential it could reuse.
What this means for you: the dashboard you sign in to never holds the keys that run your phone calls. Keep your sign-in safe, and the keys stay out of reach.
Keeping one business out of another
Each record that belongs to a business carries its organization. Access is enforced in the database with row-level security for ordinary signed-in queries, so a query scoped to one workspace does not return another workspace's data.
That enforcement is not complete yet. Row-level security is enabled but not forced on nine tables: knowledge collections, knowledge files, knowledge voice files, knowledge voice indexes, SMS threads, SMS thread messages, Studio text usage, Studio voice tool usage, and Studio voice usage. On those tables, privileged database roles can bypass the check. We are completing that enforcement, and we will update this page when it is done.
- Owners manage billing, membership and workspace settings.
- Managers work with day-to-day settings and reporting.
- Staff work with orders, bookings, callbacks and conversations.
- Privileged service operations run server-side only and are used for provisioning or billing reconciliation, not for ordinary user reads.
Calls, orders, reservations and messages created by the assistant are written under the workspace that owns the number. For ordinary signed-in use, a workspace cannot read another workspace's conversations by guessing an identifier.
What this means for you: roles limit what each person on your team can see and change. Give each person the lowest role that lets them do their job, and remove access when someone leaves.
Encryption and secrets
- Traffic between your browser, the dashboard and our backend uses HTTPS, including the real-time voice stream.
- Stored data is encrypted at rest by our database provider. We have not confirmed the same detail for each other store, so we do not claim it for them here.
- Passwords are hashed; we do not store them in readable form.
- Carrier, model and payment secrets live in server-side environment configuration and are never committed to the repository or sent to a client.
- API keys are created per integration so one can be revoked without breaking the others. API keys are stored as hashes, not in readable form.
- Dashboard requests that change data use CSRF protection or same-origin checks.
What this means for you: use a unique password for Callaroo, and do not share API keys in chat or email. If a key may have leaked, revoke it and create a new one.
Accounts and sign-in
Passwords are hashed; we do not store them in readable form. Sign-in sessions expire after 90 days, or after 180 days if you chose remember me. Activity refreshes the 90-day session.
We do not promise multi-factor authentication as a fresh check today. Admin pages check whether multi-factor authentication is enrolled, not whether it was verified recently. The enrollment flag is checked. The time of the last verification is recorded but is never compared to a time window. A session stolen long after sign-in can therefore pass the admin check.
What this means for you: treat a signed-in session as valuable for its whole lifetime. Sign out on shared devices, keep the email on the account secure, and tell us at once if you think a session or a password was taken.
Payments and card data
Subscriptions and invoices are handled by Stripe. Card numbers and security codes go to Stripe directly; they are not stored in our database and do not pass through our application servers. We keep plan, invoice and payment-status information, and only what is needed to show you the account.
The assistant is instructed not to ask a caller for card details. If a caller offers them, the intended behaviour is to decline and explain how payment is actually taken. This is an instruction to the assistant, not a guarantee. Owner-written instructions can override it today, because the check on owner text has known holes. See the section 'What we know is not finished yet'.
What this means for you: never read a card number to the assistant, and never type one into a call note, a knowledge file or a reservation note. Take payment through Stripe, not over the phone.
Telephony and messaging
- Incoming carrier webhooks are verified before they are acted on; unsigned or mismatched requests are rejected. Telnyx webhooks are checked with an Ed25519 signature and a timestamp, with a five-minute window as replay protection and event-id deduplication as a second layer.
- Stripe webhooks are checked against their signature, and each webhook is recorded so a replayed copy is rejected. If the replay check cannot run, the webhook is not processed.
- Call state changes are recorded against the provider's call identifier so a call can be reconciled without trusting a client claim.
- Internal service routes require a shared secret and are not exposed to the browser.
- Business messaging requires a carrier-approved number. Opt-outs are recorded and honoured before a send is attempted. A STOP reply blocks later sends until the customer sends START.
- Caller confirmation texts end with 'Reply STOP to opt out.' Owner notification texts and free-text dashboard messages do not carry that line yet. There is no HELP reply yet, and no quiet-hours window yet, so a confirmation text can arrive late at night.
- A queued or accepted message is not treated as delivered; the carrier's reported status is stored separately.
What this means for you: if a caller says they asked to stop messages, check the opt-out list before sending again. A STOP reply is honoured before any later send is attempted.
Logging and monitoring
We log authentication events, administrative changes and errors so a problem can be traced. Logs are written to avoid passwords, session tokens, one-time sign-in codes and full card details.
Error reports pass through a scrub step that removes personal information such as phone numbers, email addresses and card numbers before they leave our systems. We do not put call content in error reports.
Product analytics describe how the dashboard is used, and session replay records dashboard sessions with inputs masked. Text outside form inputs is not masked, so replay can show names, numbers and order details visible on screen. Replay is kept separate from call transcripts. The full description is on our Cookies page.
What this means for you: do not paste passwords, tokens or card numbers into a support message or a call note. If you report a problem, include steps to reproduce and the workspace affected, and nothing sensitive.
If something goes wrong
If we become aware of a security incident affecting business data, we investigate, contain it, and notify affected workspace owners and any authority the law requires. We tell you what we know, what we are doing, and what you should do, even while the investigation is still open.
For New York residents, notice goes out within 30 days of discovery. We also notify the New York Attorney General, the New York Department of State, the New York Division of State Police, and the New York Department of Financial Services. If more than 5,000 New York residents are affected, we also notify the consumer reporting agencies. If more than 500 California residents are affected, we notify the California Attorney General. Where data protection law requires it, we also notify the relevant supervisory authority within 72 hours of awareness.
To report a suspected vulnerability or a compromised account, contact [PRIVACY CONTACT]. Include steps to reproduce and the workspace affected. Do not test against another business's data, and do not run disruptive tests against production.
We are a young product. We do not currently hold a published third-party security certification, and this page describes the controls in the code today rather than an audit result. When a certification or independent assessment is completed, it will be published here.
What this means for you: if you suspect a problem, tell us at once through [PRIVACY CONTACT]. Do not wait for proof. We would rather hear about a false alarm early than learn about a real incident late.
What we know is not finished yet
This section lists the gaps we know about. We publish them here so no one has to guess what this page leaves out.
- Row-level security is enabled but not forced on nine tables, so privileged database roles can bypass it. See 'Keeping one business out of another'.
- The admin sign-in check looks at multi-factor enrollment, not at a fresh verification. See 'Accounts and sign-in'.
- Owner text is checked when agent settings are saved, but the check has known holes. Instructions to take payment by phone, to tell callers to buy gift cards, to claim to work for another business, or to claim to be a real person can pass it. The assistant prompt itself tells it not to claim to be human and not to take payment details, but the save-time check does not block those instructions.
- Knowledge base uploads are not scanned. PDF and DOCX files are checked for file type only. Their contents are never extracted or inspected, so text inside a document can reach the assistant.
- Reservation names and notes are not checked for sensitive numbers. A card number typed as a booking name or in notes is stored and can be read back to staff. Order and callback paths reject such input. Reservations do not yet.
- Stored transcripts and summaries pass through a number-redaction step before they are saved. That step removes digit runs from stored text only. It does not remove numbers from audio, and it does not catch numbers spelled out in words.
What this means for you: review your owner text and your knowledge files before they go live, and check reservations the same way you check orders. Do not put card numbers, passwords or tokens in any of them.
What you should do
- Give each staff member their own account and the lowest role that lets them do their job.
- Remove access when someone leaves, and revoke API keys you no longer use.
- Use a unique password and keep the account email secure.
- Sign out on shared devices.
- Do not paste card numbers, passwords or tokens into a call note or a support email.
- Review orders and bookings before service, especially in the first weeks after switching a number.
Contact
Callaroo is operated by [LEGAL ENTITY], [ADDRESS]. Questions about this page, and reports of a suspected vulnerability or a compromised account, go to [PRIVACY CONTACT].
Do not use an address that ends in .example. It cannot receive mail. Use [PRIVACY CONTACT] instead.
What changed
- 13 November 2026: Full rewrite. Added the short version, accounts and sign-in, known gaps, contact details and this change log. Softened the tenant isolation claim to match the database. Added breach notice times for New York and California. Replaced the old support address with [PRIVACY CONTACT].
- 1 November 2026: Previous version.