Security

An AI answering your phone is a real risk.

So here is how it is built, in language you can check against what you see in the product.

Your data is separated at the database level

Every record belongs to exactly one workspace, and the database itself enforces that with row-level security — not application code that could forget a filter. The application's database role cannot bypass those policies, and the workspace identity is set per transaction so a shared connection pool cannot leak it between customers. An automated test attempts a cross-workspace read, update, delete and insert on every table before any change can be merged.

Credentials are encrypted with a key per workspace

When you connect Google Calendar or a phone number, the tokens are sealed with a data key unique to that secret, which is itself wrapped by a key held in a managed key service. Your workspace identity is bound into the encryption, so a ciphertext moved to another workspace's row fails to decrypt rather than quietly working. Secrets are never returned by any API and never written to a log.

The AI can only do what it has been given tools for

Agents cannot query the database, run commands or call arbitrary URLs. Every capability is a named tool with a checked input schema and a risk level, and the allowed list is re-checked on the server every time — a model asking for a tool is not authority to run it. High-risk actions like messaging your whole list, deleting records or anything involving money always stop for a person to approve.

Guardrails are code, not instructions

Identifying itself as AI, refusing medical, legal and financial advice, refusing to act on protected characteristics, honouring an opt-out, escalating an emergency — all of these run as checks around the model, not as sentences inside a prompt. A jailbreak that changes the model's behaviour still cannot get past them.

Everything consequential is written down

Who did what, to what, when, and what it looked like before and after. Audit records are append-only: the application can add to them and read them, and cannot change or remove them.

Uploads are opened somewhere they cannot hurt you

Documents you upload are parsed in an isolated process with no database access and no outbound network. File type is checked by content, not by the extension. Web pages we fetch on your behalf resolve their address first and refuse private, loopback and cloud-metadata addresses, including through redirects.

You can take your data and leave

Export contacts, conversations, documents, configuration and audit records at any time. Deletion is a workflow with a receipt that states when backups holding the data age out, rather than a promise that it is gone the moment you click.

What we do not claim

No certifications, yet

We are not SOC 2 certified and we do not say we are. The controls a SOC 2 audit looks for — access control, audit logging, change management, encryption, tested backups, vulnerability management, incident response — are built. The audit itself is not done, and until it is, a badge here would be a lie.

HIPAA is a configuration, not a claim

Healthcare workflows are designed with HIPAA requirements in mind and stay switched off until a business associate agreement is signed and the compliance setup is complete. We will not tell you we are “HIPAA compliant” because that is a description of an organisation, not a feature list.

Reporting a vulnerability

Email security@techmavs.com. We acknowledge within two business days and triage within five. We will not take legal action against good-faith research that respects customer data and does not degrade the service.