Security and privacy
What Yi Cloud stores and what it doesn't, how keys and sign-in are protected, and how sandboxes and VMs are kept apart.
What is stored
| Stored | Not stored |
|---|---|
| Your account's email and name, and your workspace | Prompts and model outputs |
| Projects and keys, keys only as a hash | The full API key |
| For each model request: model, endpoint, token counts, cost, status and latency, for 90 days | A deployment's raw error text |
| Usage totals per day, the ledger and receipts | Command output and file contents of sandboxes |
| Sandboxes and VMs: size, state, times, prices, SSH public keys and allowed address ranges | A VM's serial output, which is read live |
| Which public address each resource used and when, and network security records |
Our logs record a request's ID, method, path, status and duration. They don't record request or response bodies.
Model API
- Prompts and outputs pass through and aren't saved. Yi Cloud reads a stream only to count tokens.
storeis always sent asfalse. Background responses, stored conversations, stored files and hosted tools are refused, so nothing is kept for you at the model either.- Requests go only to model deployments we configured, over HTTPS. Redirects are never followed.
- Error messages from a deployment are replaced with our own. Its raw text never reaches you or our logs.
API keys
- A key is shown once, when it's created. Yi Cloud stores a SHA-256 hash of it, compares in constant time, and can't show it again.
- A revoked key stops working at once.
- Each key can be limited to the uses it needs, a spend limit and a requests-per-minute limit. See API keys and limits.
Console sign-in
- Sign-in uses OpenID Connect with your Yi Labs account, with state, nonce and PKCE checks.
- The browser cookie holds only a random handle. It is HTTP-only, and Secure over HTTPS. Your identity and the account's token stay on the server, the token encrypted.
- A sign-in lasts at most 8 hours. It is rechecked with your Yi Labs account every 60 seconds. If the check fails, access stops.
- Changes must come from the console's own origin.
- The console loads nothing from other sites.
- Sign-in attempts are rate-limited per IP address. The address is kept only as a hash.
Workspaces
- Your workspace comes from your signed identity at sign-in, or from the API key. No request parameter can switch it.
- Every key, project, sandbox, VM, receipt and ledger entry is checked against your workspace.
- Changes such as creating and revoking keys and acting on sandboxes and VMs are written to an audit log.
Sandboxes and VMs
- One machine each. Every sandbox and every VM is its own virtual machine.
- No cloud identity. No cloud account or service identity is attached to a sandbox or VM, so code inside can't use one.
- Disk destroyed on delete. When a sandbox or VM is deleted, its disk is deleted with it. Reinstalling a VM erases its system disk.
- No inbound to sandboxes. A sandbox has no public address. You reach it only through the API.
- SSH to VMs only from your ranges. A VM accepts only TCP port 22, only from the ranges you listed, only with the keys you gave. Password login and direct root login are off.
- Limited outbound. HTTP, HTTPS and the platform's DNS only. Mail ports, private address ranges and known threat destinations are always blocked. See Quotas and network.
How commands reach a sandbox
- The sandbox connects out to Yi Cloud over an encrypted WebSocket. Nothing connects in.
- It proves who it is with a one-time credential that expires after 10 minutes. Yi Cloud stores only hashes of that credential and of the session that follows, and a session works for one sandbox only.
- The command daemon inside listens only on the machine's loopback interface.
- Commands run as an ordinary user without
sudo. Only/workspace, its home directory and/tmpare writable. - Command output and file contents stream through and aren't saved in our database or logs.
Inside a VM
You're root inside your VM. What you install and run there, and who you give access to, is yours to secure. Yi Cloud controls what reaches the VM from outside and what it may send out.
Abuse handling
When we suspect abuse of a sandbox or VM, we can close off its network, keep a snapshot of its disk for review, and stop or delete it. A resource closed off this way shows isolated: true and accepts only stop and delete. See Acceptable use.
Money
Every call and resource holds its cost before it starts and is never charged more than that hold. Ledger entries can only be added, never changed or deleted. Receipts come from the same records.
Reporting a problem
Use the contact form on our website. Say what you found and how to reproduce it.