Security

How Zytheris handles models, prompts, and evidence.

No customer weight uploads

There is no weight-upload form. Shelf packs are trained in our lab on public checkpoints. For a custom pack we pull a checkpoint we are granted (public ID, gated repo, or private access), on GPUs we provision for that model’s size, then import features plus evidence into your workspace. Checkpoint files never pass through the website.

What we store

  • Account and team membership
  • Evidence reports we ship with features
  • Prompts and outputs if you try a feature in the app

Offline training

Feature dictionaries are trained offline, not in the browser. Activation dumps are not uploaded through the web UI.

API keys

Account API keys can call pack, steer, and report endpoints for your team. Steer jobs run on private GPU workers. There is no public, unauthenticated inference endpoint.

Security questions: hello@zytheris.ca.