Security & trust
Safe by design, private by default.
MagizAI keeps each workspace's data apart, encrypts stored keys and credentials, and sets limits in code on what the AI can see and do, on top of the rules it is given. Here's how we keep your data and your customers safe.
Encrypted credentials
Every connector credential and AI key is stored encrypted, kept per workspace, and never sent back by the dashboard or the API once saved.
Bring your own key
Use your own AI provider keys so prompts and data flow through your account, not a shared one.
SSRF-hardened crawling
Training fetches only reach public addresses: each resolved IP is checked against private and reserved ranges, the connection is pinned to that IP, every redirect is checked again, and each page has a size cap.
Guardrails in every reply
The assistant is told to answer only from your own content, to say plainly when it does not know instead of inventing a fact or a price, and to stay on your topics. When an AI provider fails during a chat, a backup provider answers or the visitor is told so.
Multi-tenant isolation
Every query is scoped to one workspace, and a suspended workspace is locked out of the dashboard, the API and the portal, while its chatbot stops answering visitors.
Metered & auditable
Every AI call a workspace makes is recorded with its tokens and estimated cost, and sensitive dashboard actions, such as key, role and connector changes and visitor erasures, are written to an audit log.
Card numbers masked before saving
Visitors paste card numbers into chat whatever you ask them not to. On your website, WhatsApp, email and help desk tickets, a card number (Luhn-checked, in Western or Arabic-Indic digits) is replaced by its last four digits before the message is saved, so the conversation, the inbox, transcript emails and your AI provider only see those four.
Defended against poisoned pages
A page you train on could contain instructions aimed at your bot. Wherever retrieved or imported text reaches the AI, from trained pages to connector results and imported chats, common instruction phrases are removed first. In chat replies that content is also fenced off as reference data, and the assistant is told never to follow it as an instruction.
Data requests, handled
Export what is held about one visitor's chats (transcripts, contact details, browsing profile, remembered facts), or erase it together with what their chats led to: bookings and their files, quotes, reviews and help desk tickets. Records are deleted, not anonymised, because a transcript still names the person; a queue ticket keeps its number without the name and phone. Every erasure is audit-logged. Customer records are erased in the Customers app.
Accessible to everyone
The widget is keyboard-navigable, screen-reader-labelled and closes on Escape, with visible focus throughout. Chat widgets are a common target of accessibility complaints; ours is built not to be one.
The Auto-Pilot promise
Powerful actions, hard limits.
- Never types into a field it recognises as a password, a one-time code or card details.
- Asks before a payment, deletion or cancellation it recognises, and before submitting a form the visitor did not ask it to send.
- Answers how-to questions from your knowledge base and points only at what is on the page.
- Client-side redaction of sensitive fields.
Questions about security?
Our team is happy to walk through data handling, hosting and compliance for your rollout.