The security questions worth asking a chatbot vendor

A website chatbot sits on every page of your site, reads what visitors type and sends it to somebody else's server. That deserves eleven questions, and a vendor who cannot answer them in plain words has told you something. Below is each question, why it matters, and the answer for Allinaa - including where the answer is "no" or "not yet".

1. What exactly does the script on my pages read?

A chat widget can be written to read anything on the page. Ask what it actually sends. Allinaa's widget sends the address of the page and its visible text so the assistant can answer about what the visitor is looking at, and it skips private path segments such as checkout and account pages. It does not read form fields, it does not read other tabs, and it is not an analytics script. See page-aware answers.

2. Who can read what my visitors typed?

This is the question most vendors answer loosely. In Allinaa, conversations belong to your business: your own team accounts see them, and Allinaa staff do not - not even a platform administrator - unless you open a time-limited support grant for a stated reason. Grants expire, and every read under one is written to an access log you can see. Metadata such as counts is visible to Allinaa staff for running the service; the words are not.

3. Is my data used to train a model?

Ask it in those words. For Allinaa, no: your conversations and your approved knowledge are used to answer your visitors, not to train anything. The assistant answers from the knowledge you approve, and the model provider is used as a service, under the terms of the plan Allinaa pays for.

4. Where is the data, and who is the processor?

You should be able to name the countries and the subprocessors. Allinaa runs on its own server and uses a model provider to generate replies and an e-mail provider to send notifications; which, and what each receives, is in the privacy policy. If your own customers ask you this - and in Europe they will - you need an answer you can show them.

5. How long is it kept, and can I delete it?

Ask for the retention period of conversations and of files, and ask what deletion actually means. Allinaa keeps conversations for your own panel and reports, and files sent by visitors are deleted on a retention schedule rather than kept for ever. See how long a chatbot keeps things.

6. What happens to a file a visitor uploads?

Uploads are the sharpest edge of any chat product, because a visitor can send anything. On Allinaa Premium each file has its type checked against its actual content rather than its name, is size-limited and quota-limited, is checked for malware before anybody can open it, is stored under a random name outside the web root and is served only to your signed-in team with headers that stop the browser executing it. Executables, macro-enabled documents and archives are refused outright. Allinaa is not an antivirus product, and that claim is not made; what is claimed is listed. See files customers send.

7. Can the chatbot be made to say something that embarrasses me?

Any assistant built on a language model can be pushed. Ask what is deterministic rather than hopeful. In Allinaa the assistant answers from your approved knowledge; a reply that names a product price or a link is checked against your catalogue before it is sent, and a reply that claims a booking or a resolved issue is checked against the record and corrected if the record disagrees. Those are code, not instructions to a model. What remains possible is a wrong or vague answer, which is why there is a correction workflow. See how accurate a chatbot is.

8. Who on my side can change what it says?

Ask about roles and about the audit trail. Allinaa Premium has named team accounts with roles - owner, administrator, operator - for up to 5 people, every management action is written to an audit trail with who did it and when, and the knowledge is versioned, so a change is a new version you can see rather than an edit that vanishes. A chatbot that any staff member can silently retrain is a risk to your own business.

9. Does it set cookies, and do I need a banner?

Allinaa's chat widget does not set a tracking cookie; it uses a server-issued session so a conversation survives a page change. Whether your site needs a banner depends on everything else on your site, and no vendor can answer that for you - but the chatbot should not be the reason you need one.

10. What is the honest list of what you are NOT?

This is the most useful question in the list, and the answers for Allinaa are in writing: it is not HIPAA compliant and makes no compliance or certification claim; it does not know the current date or time; it will not give medical, legal or financial advice; several channels are built but waiting on platform approval, so only the website chat and Telegram are live-verified today; and it can be wrong, which is why your business approves the knowledge. A vendor who has no such list has not thought about it. See what a chatbot cannot do.

11. What happens when it breaks or runs out?

Ask what the visitor sees when the service is down, when a reply allowance is used up, or when the model provider refuses. On Allinaa the chat falls back to the inquiry form so contact capture keeps working, and the business is notified rather than finding out from a customer. Ask every vendor this: the answer tells you whether they have watched their own product fail.

Questions people ask about this

Should I ask for a penetration test report?

Ask, but judge the answer rather than the document. A small vendor who explains precisely how uploads are handled and what staff can read is worth more than a years-old report about a different product version.

Is a self-hosted chatbot safer?

It moves the risk rather than removing it: you gain control of the data and take on the patching, the backups and the monitoring. If nobody on your side will do those weekly, hosted is usually safer in practice.

What is the minimum I should insist on?

Written answers to questions 2, 3, 5 and 10. Who can read it, whether it trains a model, how long it is kept, and what the product is not.