Sees the page, not your users’ data
This page describes exactly what the guide reads, what it never reads, and where it runs. If something here doesn’t answer your security team, write to [email protected] and we will add it.
What the guide reads
- The page path, without the query string, and the page title.
- Headings, to understand where the user is.
- Visible interactive elements: their role (button, link, field), their label, and which region they sit in.
- The user’s goal, as typed, and the steps taken so far in this goal.
What it never reads
- Values typed into inputs, text areas and selects.
- The HTML of the page, cookies, local storage or screenshots.
- Anything inside
data-guide-ignore. - Cross-origin iframes and closed shadow roots, which the browser keeps private anyway.
PII masking happens in the browser
Before a request leaves the page, labels are scrubbed: email addresses and long digit sequences (account numbers, phone numbers, card-like numbers) are replaced. The text of any element inside data-guide-private becomes “(private)”.
The server then keeps the guide honest: if the brain answers with an element id that isn’t in the snapshot, the step is turned into a plain message. A hostile label on a page can’t make Nuri point outside that page.
<!-- The guide sees "(private)" instead of the customer's name -->
<section data-guide-private>
<h2>Acme Ltd, account 4419…</h2>
</section>
<!-- The guide does not see this at all -->
<div data-guide-ignore>…</div>Where it runs
Hosted
The widget calls the Wayhint API, which asks the model for the next step: DeepSeek V4 Flash, an open-weight model served by DeepInfra. Prompts hold your site description, passages from your help center, the user’s question and the masked snapshot.
Your own model
Per site, in the dashboard: any OpenAI-compatible endpoint (your own vLLM or Ollama behind https, or a provider you already use) or an Anthropic key, stored encrypted. Guide requests for that site then go to your model, not ours.
Your infrastructure (planned)
Planned for Business: run the guide API in your own infrastructure. The widget already supports pointing at another API with one attribute, data-api.
Controls for site owners
- Origin allowlist: a site key only works on the domains you list.
- Rate limits per route on the API.
- Site keys are public identifiers; there are no secrets in the widget.
- The widget renders in a shadow root and only adds one element to your page.
- Nuri never acts for the user. It points; the user clicks.
- If our service is down, the widget hides itself and your app works as before.
- Restricted areas: list the parts of your app only some roles use, and Nuri never points into them for anyone else.
Observe mode and the app map
Observe mode is off unless you turn it on. When on, it records the masked page structure and which control led where, with a random id per browser tab and no user identifier. A label is used only after several separate sessions have seen it, so one user’s data on a button never reaches the prompt. Records are kept 90 days by default.
Account tracking
Also off unless you turn it on for a site. Your page then passes its own opaque id for the customer account (and an optional display name), with a one-way hash of the user id. Ids that look like emails, phone or bank numbers, or secrets are refused in the browser and again on the server. The account id and name are never sent to the model.
Team roles
Each account has owners and members, and the dashboard checks the role on every change, not only by hiding buttons.
- Only owners change the allowed origins and the support URL, choose the AI model and its key, set up alerts and webhooks, delete a site, and invite or remove people.
- Invites are one-time links that expire after 7 days. Removing someone signs them out everywhere at once.
- A password change or reset signs out every other session.
Secrets, encrypted
The secrets you give the dashboard are encrypted with AES-256-GCM before they are stored, and never shown again:
- your own model’s API key;
- Slack and webhook URLs, and webhook signing secrets;
- help-center API tokens (Intercom, Freshdesk, Help Scout).
The dashboard shows only a host or the last 4 characters. The encryption key can be rotated.
Signed webhooks, safe alerts
Every webhook request carries a Wayhint-Signature header: an HMAC-SHA256 of a timestamp and the raw body, keyed with your signing secret. Check it in constant time and refuse timestamps more than 5 minutes old; the dashboard’s help has code for Node.js and Python. A new secret overlaps the old one for 24 hours, so you can rotate without losing events.
- Alert and webhook URLs must be https on a public address. Redirects are not followed.
- Questions and comments in alerts and events are masked. No event identifies a user, not even by hash.
- Preview conversations never trigger an alert.
- Rate limits per site keep a flood of events from crowding out a visitor asking for support.
Sub-processors
The companies that handle data for the hosted guide today:
| Company | What for | What it receives |
|---|---|---|
| DeepInfra (US) | Runs the hosted model, DeepSeek V4 Flash | For each step: the masked page snapshot, the user’s question, your site description and help-center passages. Per DeepInfra’s data privacy page (checked 8 October 2026), inputs are not stored to disk during inference and not used for training. |
| Cloudflare | Serves this website and its demo; DNS; carries traffic to our API and dashboard (Cloudflare Tunnel) | Requests in transit, including the widget’s calls to the API |
| Your model provider | Only if you choose your own model for a site | The same as DeepInfra above, under your own terms with that provider |
The API, the dashboard and the database run on a machine we operate; no cloud hosting provider stores your data today. We will update this list before adding a sub-processor.
Two flows don’t go through a sub-processor of ours. The support handoff summary goes from the visitor’s browser straight to the support tool on your page (Intercom, Zendesk, Crisp and others), WhatsApp or email. Alerts and webhooks go only to the endpoints you set, under your terms with those tools.
Retention
- Guide analytics (questions asked, steps served, outcomes, nudge and checklist counters): deleted after 180 days.
- Account data, when account tracking is on: an account’s conversations go with the guide analytics after 180 days, and the account row and its checklist progress once the account hasn’t been seen for 180 days. One account’s data can be deleted at once from its page.
- Deleting a site deletes its guide analytics and account data at once, in the same step.
- Alert and webhook delivery logs: 7 days for delivered events, 30 days for the others.
- Observe-mode records: deleted after 90 days.
- Your site settings, help-center passages and flows: kept until you delete them or the site.
GDPR and compliance status
Wayhint is designed to process as little personal data as possible: no input values, masked labels, no user identifiers in observe mode. Where personal data may still appear in page labels, you control it with the attributes above.
We are in early access. We do not hold SOC 2 or ISO 27001 certification today, and we will say so until we do. A one-page DPA is in preparation for our design partners; region choice is planned for the Business plan. Our privacy policy is a draft. Questions: [email protected].
Strict Content Security Policy and Trusted Types
The widget works under a strict policy: no 'unsafe-inline' for scripts or styles. Add the widget’s origin to script-src, connect-src and style-src.
With Trusted Types enforced, add wayhint to your trusted-types list, next to the policy names your own app and libraries already use (for example trusted-types myapp wayhint). The widget builds its interface without HTML strings; its wayhint policy only loads the widget’s own chunks and refuses any address outside the folder the widget came from. With a nonce-based policy, put the nonce on the script tag and the widget passes it on.
A custom 3D character from a model file also needs blob: in img-src and connect-src. More in the install guide.
default-src 'self';
script-src 'self' https://api.wayhint.com;
style-src 'self' https://api.wayhint.com;
connect-src 'self' https://api.wayhint.com;
require-trusted-types-for 'script';
trusted-types wayhint;Security questions
Does Wayhint see what my users type?
Not in your app. The widget never reads the values of inputs, text areas or selects; it only reads what a control is called, such as "Email" or "Save". Nuri never reads form fields or what users type into your app. The question a user asks Nuri is sent to the model, with emails and long numbers masked.
What exactly is sent to the server on each step?
The page URL without the query string, the page title, headings, and a list of interactive elements with a short id, a role, a label and the page region. Emails and long numbers in labels are masked in the browser before the request is made.
Can we run it without sending data to a third-party model?
Yes, per site: in the dashboard, point the site at your own model, either an OpenAI-compatible endpoint (your own vLLM or Ollama behind https, or a provider you already use) or an Anthropic key. Guide requests then go from our API to your model instead of DeepInfra. Running the guide API itself in your infrastructure is planned for Business.
Is Wayhint SOC 2 or ISO 27001 certified?
Not yet. Wayhint is in early access and we will not claim certifications we do not have. We can answer a security questionnaire and share the architecture in detail.
Where is data processed and how long is it kept?
The widget sends masked page structure and the user’s question to our API, which runs on a machine we operate, with its database. To choose each step, the API sends them to the model: DeepInfra (DeepSeek V4 Flash) or your own. Guide analytics (questions, steps, outcomes) are deleted after 180 days and observe-mode records after 90 days. A one-page DPA is in preparation for design partners; region choice is planned for Business.
Let Nuri take it from here.
Free for 3 months for design partners. We set it up with you; you paste one script tag.