Skip to content

Security overview ​

Yi Code runs an AI agent on your code, so the agent is treated as untrusted. It works in a sandbox with no network, as a non-root user, and never holds a credential. Everything it produces comes back as a diff that a person reviews before anything is published.

The sandbox ​

  • gVisor. Each turn's commands run in an OpenSandbox sandbox under gVisor (runsc). A host on macOS or ARM, or one that falls back to the ordinary runc runtime, is refused.
  • No outbound network. Sandboxes sit on a dedicated internal Docker network with IPv6 and container-to-container traffic off. The host's firewall drops every connection a sandbox opens. Only replies for the native tool transport get through. Before an execution pool is enabled, its host proves that a sandbox can't reach the internet, the cloud metadata service or listening ports on the host.
  • Not root. The agent and the verification command run as the user agent (uid 1000). The starting code's Git baseline and the original tests are prepared by a trusted process, owned by root, and can't be changed by the agent. Task containers have no Docker socket.
  • The model loop is outside. The agent's reasoning runs in a trusted runner on the host, not in the sandbox. The sandbox only carries out terminal and file-editor calls. Model and tool calls go from the runner through the control plane's gateway.
  • One turn per host at a time. A fresh virtual machine is deleted after its turn. A warm host destroys every sandbox of a turn and confirms it before taking another. If cleanup can't be confirmed, the host isn't reused.
  • Limits. Each sandbox has 2 CPUs and 4 GiB of memory. Ask turns stop after 15 minutes, Code turns after 60.

Credentials ​

  • Model credentials stay in the control plane. The agent never receives them.
  • Each execution of a turn gets its own short-lived tokens, one for reporting back and a separate one for model and tool calls, which can't be used for the other. When the execution ends, is cancelled or loses contact, its tokens stop working at once.
  • No credential goes into a repository URL, a command argument, a turn's stored configuration, a public event or the sandbox's environment.
  • For reading repositories, GitHub tokens are installation tokens limited to one repository and one purpose. Publishing uses your own GitHub authorization, stored encrypted with a key only the API holds. The GitHub App's private key and secrets are used only by the API.
  • Secrets for models, Git, knowledge and MCP tools are stored as server-side environment values and referenced by name. The app never shows a value.

Logs and exports ​

  • Logs drop task and session secret fields and replace known credential values with a placeholder.
  • Before the patch, the changed files or a memory draft leave the sandbox, they are checked for the credentials the platform knows, including in decoded content. If one is found, the whole export is refused. It is never silently edited.
  • Binary patches that use delta encoding are not exported, so a removed credential can't be rebuilt from them.

This protects the credentials Yi Code itself holds. It is not a general secret scanner. Secrets already in your repository, or typed into a message, are not detected.

Ask is read-only ​

Ask turns have only view, grep and glob. They read files inside the repository only, refuse symbolic links that point outside it, and have no terminal, no editing and no MCP tools.

Untrusted input ​

Repository files, retrieved knowledge and project memory are given to the agent as task data, and the agent is told not to take them as instructions. The internal prompt that Yi Code assembles is hidden from the turn's public events.

Tenants ​

  • Your organization comes from the signed identity at sign-in. No request parameter can switch it.
  • Every session, run, file, knowledge source, memory version and credential reference is checked against your organization.
  • Each organization has its own knowledge backend account.
  • Endpoints for models, knowledge and MCP tools must be on the deployment's allowed host list. On the cloud service, that list is ours.
  • Execution pools and rates are managed by the platform operator, not by organization administrators.

Sign-in and sessions ​

  • OpenID Connect with signature, issuer, audience, nonce and PKCE checks.
  • The browser cookie holds only a random handle. Identity and the provider's token stay on the server, the token encrypted.
  • A session lasts at most 8 hours. Membership and role are rechecked with the identity provider every 60 seconds by default. Removed members lose access at the next check, and if the check fails, access stops.
  • Requests that change something must come from the same origin.

GitHub ​

  • The app is installed only on the repositories you select.
  • An installation is connected only after GitHub confirms that your account can see it. One installation belongs to one organization.
  • GitHub's webhook messages are verified by signature, and each delivery is processed once.
  • Pull requests are drafts, created and updated only after a person confirms, with that person's own GitHub authorization. Write access to the repository is checked on GitHub before every publication. Updates are ordinary fast-forward pushes, never force pushes. Yi Code never merges.

Money ​

A turn's budget is held before it starts, and a turn is never charged more than its budget. Ledger entries can only be added, never changed or deleted. Unknown model usage is never treated as zero.

Reporting a problem ​

Use the contact form on our website. Say what you found and how to reproduce it.