Secure, and still easy to reach

1. What we do against misuse

Every conversation with your digital assistant passes a number of fixed checks.

• Every message is checked. Attempts to make the assistant forget its rules, reveal its instructions or pretend to be someone else are stopped. Also when they are hidden in a code, a different spelling or an ordinary question.
• Inappropriate content does not get through. Offensive, hateful or sexual messages are recognised and stopped. The assistant stays calm and points the way to you if something is really wrong.
• Every answer is checked before it is sent. No internal instruction leaves the assistant, and it does not invent amounts, discounts or promises that are not in your knowledge base.
• Too many messages are slowed down. Anyone who sends a very large number of messages in a short time, for example an automated program, is slowed down automatically.
• Only from your own website. The chat window only works on your own domains. Nobody else can put your digital assistant on their site or call it from outside.

2. Sensitive data

• Citizen service numbers (BSN) and bank account numbers are made unreadable before a conversation is logged. If a visitor types their BSN or IBAN in the chat, the log only shows [BSN] or [IBAN].
• Conversations are deleted automatically after at most 90 days. You can set a shorter period in Mijn Chatunda. A visitor can also have their own conversation deleted straight away.
• Never someone else's data. With Order status the assistant only looks up an order if the visitor knows the order number and a second detail, such as the email address or the amount. Anyone asking for someone else's order sees nothing, whatever the excuse.
• No IP addresses in reports. We keep the IP address of a blocked message for seven days so we can stop misuse. After that only the number per day and the type remain, without IP address and without text.

3. Your own server

Every business gets its own server. Data from different customers is never on the same server, and the server is set up so that as few doors as possible are open.

• The server can only be reached through Cloudflare, which protects the traffic against attacks. Reaching it directly does not work.
• Logging in to the server is only possible with a digital key, never with a password. Anyone who tries to log in too often is blocked automatically.
• Security updates are installed automatically. If an update needs a restart, that happens at night.
• The management functions of your assistant cannot be reached from outside. We manage everything through a secure connection.
• Every night we make a backup, and an encrypted copy is kept with a second provider in Amsterdam. Only Chatunda has the key.

New servers get this setup automatically, existing servers get every improvement with the next update.

4. Secure, and still easy to reach

A digital assistant that shuts down at every question containing the word "system", "code" or "instructions" does not help your customers. An angry customer, a question full of typos or someone who wants their data deleted must simply be helped too. That is why we do not only test whether misuse is stopped, but also whether ordinary, tricky questions simply get an answer. Both have to be good.

5. Tested every month

Before you receive the demo of your digital assistant, and every month after that, we run a fixed security test. In it we try to mislead the assistant, make it leak something and make it write inappropriate content, we ask ordinary tricky questions, and we check the server. This produces two scores, each on a scale of 100:

• Resilience: how much of what should be stopped is actually stopped.
• Accessibility: how many ordinary questions simply get an answer.

Both have to be 90 or higher. If not, we improve first and test again. Without a recent, passed security test you do not get a demo from us.

6. What you see every month

The monthly report of your digital assistant has a Security chapter. In it you read:

• the result of the latest security test, and what we improved since the previous test
• which attempts at misuse there were that month: the type, the date and the time
• how often someone was slowed down for sending too many messages

You receive the full security report as a PDF attachment, with the parts of the test and the checks of your server. The text of those messages and the IP addresses are deliberately left out. That is better for privacy, and it gives an attacker no examples.

7. Who we work with

Security does not stop at our own code. Behind your digital assistant are seven parties, and for every party that sees any of your data, we ask the same five questions. Only when the answer to all five is good enough do we work with them.

• Where. Where is the data, and does it stay within the European Union? If not, on what legal basis does it leave the EU?
• Who. How is the party organised, how long has it existed, and what is its interest in our data?
• Verifiable. Which standards have been independently audited, not merely promised? ISO 27001, SOC 2, BSI C5.
• As little as possible. Does the party receive only what is strictly necessary for its task, and nothing more?
• Control. What do we keep in our own hands, and is there a data processing agreement that sets out what the party may and may not do?

A good supplier is a foundation, not a guarantee. And a server in the EU does not automatically make a service GDPR-proof: data processing is a chain. That is why below you find, per party, the considerations and the facts, and then the whole chain in one picture.

Hetzner: the server of your assistant

European data centres, independently audited security standards, full control over our own servers, room to grow and a fair price. A new server is set up within fifteen minutes, which allows us to give every customer a server of their own without hosting costs driving up your monthly fee. Not because Hetzner is the only choice, but because the combination is right.

Scaleway: the second place for your backups

A backup that sits in the same place as the original is not a backup. That is why a copy of every server goes to Scaleway, a French cloud provider with a data centre in Amsterdam: another provider, another country, another owner than Hetzner. The copies are encrypted before they leave the server, with a key only Chatunda holds. Every copy has a fixed retention period that nobody can shorten, not even us, and the key our server uploads with cannot delete anything. So a break-in on our own management server cannot wipe the backups, and nothing is ever kept longer than we promise you.

OpenAI: the AI model that answers

This is the choice where it is uncomfortable, and we would rather say so ourselves. The model that produces your assistant's answers runs at OpenAI in the United States. There is currently no European model that delivers the same quality for your visitors' questions, and quality is exactly what your assistant is judged on. So we deliberately choose OpenAI, and make sure the model sees as little as possible: every message has already been checked before it goes to the model, and what goes to OpenAI cannot be directly traced to your business or to the visitor who is chatting at that moment. If the model falters, your assistant switches to a backup model at the same party.

Anthropic: when testing and checking

Anthropic, the maker of the Claude models, never talks to your visitors. We build and maintain the service with Claude Code, and use Claude to test the quality and security of your assistant and to check the knowledge base: hundreds of test questions before the demo, the monthly security test, and the individual visitor questions your assistant did not answer well. For that work Claude delivers the most careful assessments in our measurements. Anthropic too is an American company, so the same applies as with OpenAI: passing on as little as possible, and no training on customer data.

Cloudflare: the protection of the traffic

All traffic to your assistant and to chatunda.com passes through Cloudflare's network, the largest and best-tested network of its kind. It absorbs attacks before they reach your server, hides your server's address and provides the encrypted connection. Without such a shield, a server on the internet is a target within a day. Traffic from the Netherlands is handled in Amsterdam, so within the EU. Cloudflare is an American company, but the only one of the three that is verifiably certified under the EU-US Data Privacy Framework.

Zoho: the e-mail to you

The e-mail Chatunda sends to you, such as the questionnaire invitation, the monthly report, invoices and outage notices, goes through Zoho's ZeptoMail. We use only Zoho's European environment, and a service that only sends e-mail, not a marketing platform that builds profiles. Visitors of your website never receive e-mail from us. Zoho is an Indian company with a European establishment.

Mollie: the payment

You pay through Mollie in Amsterdam: iDEAL/Wero for the set-up fee and a mandate for the monthly direct debit. Mollie is a Dutch payment service provider licensed by De Nederlandsche Bank. Your bank details and your mandate are held by Mollie, not by us. We receive only the account holder's name, the account number, the amount, the date and the payment status, because we need those for your invoice.

8. The chain in one picture

This is what happens when a visitor of your website asks a question.

1. The visitor types a question in the chat window on your website. The chat window is loaded from your own server at Hetzner.
2. The request goes through Cloudflare (Amsterdam) to your server. Cloudflare absorbs attacks and sees the visitor's IP address.
3. On your server at Hetzner (Nuremberg) the message is checked for manipulation and unwanted content, and the relevant pieces of your knowledge base are looked up.
4. The question, the last messages and those pieces of knowledge go to OpenAI's model (United States). No IP address, no session ID. The answer comes back.
5. The answer is checked once more for promises your assistant must not make, and goes back to the visitor through Cloudflare.
6. The conversation stays on your server for at most 90 days, or shorter if you set that in Mijn Chatunda, and is then deleted automatically. Every night an encrypted copy goes to Scaleway (Amsterdam), which is wiped automatically after 30 days at most. The weekly full copy after 56 days.

Anthropic, Zoho and Mollie do not appear in this conversation. Anthropic only works when testing and checking your assistant, Zoho only sends e-mail to you, Mollie only processes your payment.

If anything in this chain changes, we update this page and the privacy statement and inform our customers by e-mail.

9. If something does go wrong

If we discover a breach that affects the data of your business, we report it to you within 36 hours at the latest, with what happened, who is affected and what we are doing about it. That leaves you time for your own report to the Dutch Data Protection Authority. This is also in article 9 of the data processing agreement. If you notice something odd about your digital assistant yourself, or have a question about this page, email info@chatunda.com. You will hear from us within one working day.