Security & limitations

We are not going to tell you VerifAI catches 100% of deepfakes. Nobody credibly can. Here is what we do, and what you should still assume you need to do.

Data handling

  • Every uploaded video is stored in a private, org-scoped object storage bucket. Only signed URLs are ever served.
  • Row-level security scopes every database read to the requester's organization. Cross-tenant reads are impossible at the database layer.
  • API keys are hashed at rest. The full key value is shown once at creation and never again.
  • All privileged actions (verdict overrides, key creation, membership changes) are written to an immutable audit log.

What our detectors do well

  • Catch commodity face-swap and lip-sync tools - most published deepfake generators leave detectable frame or temporal artifacts.
  • Flag re-encoded or metadata-stripped video, which is a strong prior for tampering in KYC contexts.
  • Verify signed content credentials (C2PA) when present.

What we can't guarantee

  • A state-of-the-art bespoke generator, given enough compute and no time pressure, can defeat any single detector we run - including ours. That's why we run six.
  • Very short (<3s) or very low-quality clips give any detector less signal to work with. Confidence bands reflect this.
  • Live-stream detection is out of scope in v1; VerifAI targets pre-recorded video.

How to use it well

Treat VerifAI as one control in a defense-in-depth stack: pair it with liveness challenges, device signals, transaction risk scoring, and - for anything high-value - a trained human reviewer looking at the evidence page we produce.