Security
Avenue Lift runs a script on your website and stores data about your experiments. This page describes how each of those is kept safe, in terms specific enough to check.
Last updated
The snippet on your site
Installing a third-party script is a matter of trust, so the snippet is built around one rule: it must never break the page it runs on.
- It fails to the original page. Every step runs inside error handling. If the configuration cannot be loaded, or anything throws, the visitor sees your page exactly as you built it.
- The anti-flicker hold has a hard time limit. The page is briefly held back while changes apply so visitors never see it switch. If the limit passes, the page is revealed regardless. You can shorten the limit or turn the hold off per site.
- It is small and has no dependencies. One file, served from the same application as its configuration. It loads nothing from other third parties.
- You can switch it off for a page view. Add
?lift_off=1to any URL to load the page without it, which is useful when you are debugging your own site.
What a variant is allowed to do
A variant is not code. It is a list of declarative changes from a fixed vocabulary: replace text or HTML, set an attribute or style, add or remove a class, hide, insert or move an element, add CSS, or redirect to another URL.
- No arbitrary JavaScript, ever. There is no change type that executes script, whether the variant was written by a person or by the AI.
- HTML is sanitized twice. Once on the server when a variant is saved, and again in the browser, where it is inserted through an allow-list of elements and attributes.
- AI output is a draft by default. An AI-written variant for a running test is held for review and is not served to anyone until a member of your workspace approves it. The one exception is a setting an admin must turn on per site, auto-approve AI variants: on that site AI-written variants are served without review, still limited to the change types above.
Keeping workspaces apart
Avenue Lift is multi-tenant: many organizations share one application and one database. Isolation is enforced in the layer that reads and writes data. Every operation there takes the caller's organization and scopes its queries by it, and no other part of the application talks to the database directly.
The public keys in your script tag identify a site and can only do what the snippet needs: fetch that site's running experiments and report events for it.
Sign-in, roles and the audit log
- Sign-in is handled by WorkOS. We do not store passwords.
- Roles limit what each member can do. Viewers read results. Members create and edit experiments. Admins manage sites, members, API keys and webhooks. Only the owner manages billing.
- Client report links are read-only and show one site's results; they grant no access to the workspace.
- The audit log records changes made in the workspace, by people, API keys and automation, with who and when.
API keys and webhooks
- API keys are stored as hashes. A secret key (
lift_sk_…) is shown once when it is created. We keep only its SHA-256 hash, so it cannot be read back out of our database, by us or anyone else. Keys carry scopes and can be revoked at any time. - Webhooks are signed. Each delivery carries a timestamped HMAC signature made with the endpoint's secret, so your server can verify that it came from us and reject replays. Secrets can be rotated; the old one stops working immediately.
Visitor data
The safest data is the data that was never collected. The snippet records a random visitor ID, the variant shown and the conversions on your goals. It does not collect names, emails or form contents, and the application does not read, store or log IP addresses. Browsers sending Do Not Track or Global Privacy Control are left out by default.
The full list, including retention, is on the privacy page.
The AI and outbound requests
- Reading your pages. To propose tests, the application fetches pages of your site over the public internet. Those requests pass through a guard that refuses private, loopback and internal network addresses, so a crafted URL cannot be used to reach systems behind our network.
- What the model sees. Page content, your brand notes, earlier learnings and aggregate results. Not individual visitor records.
- Checked numbers. When the AI writes a summary of results, the figures in it are checked against the ones the statistics engine computed.
Secrets, payments and infrastructure
- Secrets such as provider credentials live in the server environment. They are never sent to the browser and never written to logs.
- Payments are handled by Stripe. Card numbers are entered on Stripe's pages and never reach our servers.
- Hosting. The application and its PostgreSQL database run on Railway. The other providers we rely on are listed on the privacy page.
- Abuse controls. Event ingestion is rate limited and filters automated traffic, so a flood of requests cannot fill your results with noise.