What leaves the device, and what doesn’t.
An on-device renderer has a genuinely smaller attack surface than a streaming one. It is not zero, and this page is the honest version.
Data handling
The Yoob SDK transmits no audio, video, transcripts or rendered frames to Yoob. Rendering and speech synthesis happen on the user’s device, and there is no render pipeline off the device for a frame to travel through. A streamed avatar necessarily sends the conversation to a vendor to be rendered; this one structurally cannot.
Where the boundary sits, stated plainly: this is not offline AI. In the SDK you bring your own conversation model, and whatever you send it goes to your provider under your key — we never see it. In the Companion, the conversation model is ours and runs on our GPUs, so what you say does reach us. Both facts are on this page because a claim with an undisclosed exception is worth less than no claim.
The Companion product does send what you say to our conversation backend, because that is where the model runs. We do not sell it, we do not use it to train third-party models, and you can delete it — see Privacy.
Model protection
A determined attacker can extract a Core ML model from any iOS application. Compiling it is not protection and we are not going to claim otherwise. What we do is raise the cost and make theft traceable:
- Never in the binary. Weights arrive over an attested channel after install, so a decompiled
.ipacontains no model at all. - Encrypted at rest, with a key provisioned per device and held in the Secure Enclave.
- Per-licensee fingerprinting. Each customer’s weights carry a quality-neutral perturbation, so a leaked model identifies its source. The perturbation is gated through the same quality battery, so it provably does not change output.
- The weights alone are not enough. Correct output depends on a conditioning contract we serve signed rather than compile into the client.
Enterprise controls
| Control | Availability |
|---|---|
| All network calls disabled, pre-provisioned licence | Enterprise |
| Model-weight and source escrow | Enterprise |
| On-premise character training | Enterprise |
| Data processing agreement | Studio and above |
| SSO / SAML | Enterprise |
| Signed bundles with published checksums | All plans |
What we do not have yet
We are not SOC 2 certified and we are not going to imply it. Several vendors in this market are, and if certification is a hard requirement for your procurement team today, we are not the right choice yet. It is on the roadmap; ask us where it stands.
Reporting a vulnerability
Email security@yoob.com. We aim to acknowledge within one business day. Please give us reasonable time to ship a fix before publishing.
Last updated 26 July 2026.