Trust & safety
Security at Looplift
Looplift runs experiments on your live site, so the honest question is not “is there a badge” but “what exactly can this thing do to my page, and who decides?” This page answers that specifically.
The approval model: nothing auto-publishes
The AI proposes; a human disposes. No experiment ever reaches your visitors without a person on your team explicitly approving its launch, and no winning variant becomes your new baseline without a person explicitly promoting it.
Approvals are pinned to the exact variant version that was reviewed. If a variant is edited after approval — by anyone, for any reason — the approval is invalidated and the edited version needs its own. There is no window in which an unreviewed change can ride an old approval into production.
Deterministic safety rules live in code, not in prompts. Model output is treated as untrusted input: it is schema-validated, it cannot write to our database directly, and it cannot set an experiment's status.
What the snippet can do
- Apply changes from an allow-listed set of declarative operations — edit text, hide an element, adjust styles, insert markup at a scoped selector — defined and validated in code, not written freehand by a model.
- Bucket a visitor into an A/B group with a first-party cookie and record experiment exposures and conversions.
- Serve only experiments a human explicitly approved, at the exact version they approved.
What the snippet cannot do
- Execute arbitrary AI-generated JavaScript in your visitors' browsers. Model output is a proposal that must pass schema validation; it never becomes code.
- Break your page when something goes wrong. The snippet fails open: if a patch cannot apply, or our config endpoint is unreachable, visitors see your unmodified site.
- Read form contents, collect personal information, or track visitors across other sites.
- Persist anything on its own authority. Every change is versioned, scoped to its experiment, and reversible — including promoted winners, which become versioned baselines you can roll back.
There is also a kill switch: you can pause every experiment on a site with one control, which empties the config the snippet receives.
Data & tenancy
- Tenant isolation is enforced by the database, not by application code remembering to filter. Every tenant table carries Postgres row-level security policies. We also keep an automated suite that rebuilds the schema from scratch, seeds two organizations and asserts neither can read the other's data — the check that catches a table shipped without a policy.
- EU-hosted data. The database, authentication, and stored audit artifacts live in Supabase's eu-central-2 region.
- GA4 tokens are encrypted at rest and scoped read-only to the single property you choose. Disconnecting revokes our access.
- Public report links are unguessable tokens, shared only when you publish a report, and unpublishable at will.
Subprocessors
The complete list — the same one in our privacy policy:
- Supabase — Authentication, Postgres database, and object storage. Primary database hosted in the EU (eu-central-2).
- Vercel — Application hosting and edge delivery.
- DigitalOcean — Hosts the audit worker that runs the analysis pipeline.
- Anthropic — The model that analyses audited pages and drafts variants. Your content is not used to train models.
- Firecrawl — Scrapes the sites you ask us to audit — nothing else.
- Polar — Payments and billing. We never store card numbers.
- Resend — Transactional email delivery.
- Microsoft Clarity — Product analytics inside our own app.
- Google Analytics (GA4) — Only if you connect it: read-only access to the property you choose, held as an encrypted refresh token you can revoke at any time.
Responsible disclosure
Found a vulnerability? Email support@looplift.io with “Security report” in the subject line. Reports go straight to the founder, and a real reply — not an autoresponder — comes within one business day. Please give us a reasonable window to fix an issue before disclosing it publicly; we will credit you if you want to be credited.
What we don't do
- We don't build personalization profiles of your visitors. Experiments are A/B tests on pages, not per-person targeting.
- We don't sell or share your data with anyone beyond the subprocessors listed above.
- We don't require a Git connection or deploy access to your codebase — the snippet is the entire integration surface.
- We don't auto-publish. Ever. Every launch and every winner promotion is a human decision.
- We don't claim certifications we don't hold. Looplift is a young company and has not completed a SOC 2 or ISO 27001 audit; what you read on this page is enforced in code and in the database instead.