Skip to content

Security ​

Scope: current hosted production uses organization workspaces. Personal workspaces and owner self-review require a later control-plane upgrade; self-hosted installations depend on their actual version and identity configuration.

The data boundary ​

With the SDK or the MCP proxy, decisions are made in your environment. The executor keeps a verified copy of the policy and decides each action locally, before it runs. Nothing about the action has to leave your environment for Guard to decide.

What the executor sends to the control plane is a redacted event:

SentNever sent
Event, run, trace, action, project and agent IDsArgument values
The trusted subject's IDTool results
Action type, resource name, destination, data labels, enforcement pointPrompts and model output
Argument namesSource code
The decision: effect, rule, reason code, message, policy version and hashYour business credentials
Status, the type of result or error, SHA-256 digests of arguments, result and error textThe text of error messages

Before the digest is taken, values under keys that look like secrets (password, token, API key, authorization, cookie, private key) are replaced. The control plane filters events again when it receives them and keeps only the fields above.

An approval ticket carries a short redacted preview and a digest of the action, never the raw arguments. Arguments for approval resume stay in a local snapshot file on your side.

Resource names, destinations, agent IDs and subject IDs can still be sensitive. Treat events and exports like other logs.

What the hosted options handle ​

The fully hosted gateway and runs are different: there, we handle the requests and responses on the gateway's routes and your upstream credentials, or the task input, results and logs of a hosted run. They are refused unless you tick consent for each deployment or run. Upstream credentials are encrypted in your browser before upload. Run results are encrypted, downloadable through a one-time link for 5 minutes, and deleted after 7 days.

The hosted control plane ​

We run the hosted control plane as a single node with daily backups. See Hosted control plane for what that means for availability. When your organization stops using it, credentials stop working at once and your data is deleted on request. Copies in backups expire within 30 days.

Accounts and access ​

  • People sign in with a Yi Labs account, or with your enterprise SSO (OIDC) on a configured project. Guard keeps no passwords.
  • Console sessions last 15 minutes and are not extended. Session cookies are HttpOnly and SameSite. Every write checks a CSRF token and the Origin.
  • Roles are per project: owner, policy admin, reviewer, viewer. The server checks each request.
  • Separation of duties. A policy's author can't review it. Another trusted subject must. A reviewer can't approve an action whose trusted subject is themselves. An owner can't approve actions in place of a reviewer.
  • Disabling a member ends their sessions at once, in every project of the organization.

Credentials ​

CredentialScopeLifetime
Executor credentialOne agent in one project. Shown once, in the connection package90 days by default. Rotate or revoke on "Connect"; a revoked agent is cut off at its next contact
SDK download tokenDownloading the SDK only. It can't call the control plane1 to 365 days. Shown once
Project tokens (self-hosted)One role in one project. The owner token issues the others and never goes into an agentRotate and revoke by API

A revoked or expired credential is never answered from the cached policy. The executor stops deciding and the tools don't run.

The SDK supply chain ​

  • The SDK is not on public PyPI or npm. Those names are unregistered there, so a public install could fetch someone else's package. Install only from your private index.
  • Each SDK version is built once and its hashes are recorded. The install command pins the wheel's sha256. The npm lockfile records the package's sha512 integrity. A file with other bytes fails to install.
  • .npmrc sends only the @preworld scope to the private index, and holds no credential.

Network ​

  • The SDK talks to the control plane over HTTPS and ignores HTTPS_PROXY and similar settings. A proxy or private CA must be set explicitly, and applies to control-plane traffic only.
  • Guarded HTTP, the broker and MCP HTTP upstreams check every address a host name resolves to, connect only to checked addresses, and refuse private and loopback addresses unless you allow a range per origin. They don't follow redirects by default.
  • The MCP proxy's HTTP downstream checks Host and Origin and needs its own bearer token.

Limits ​

We'd rather you know these before you rely on Guard:

  • It protects what goes through it. Unwrapped tools, provider-hosted tools, shell commands and raw sockets bypass it. Use network rules, the MCP proxy in its own process and a sandbox to close those paths, and test them with a call that should fail.
  • It can't promise to stop every prompt injection. It limits what the agent can do, within the rules you publish.
  • No content inspection. It doesn't read or filter HTTPS content and has no general data loss prevention. Forbidden markers are exact byte matches.
  • Pins prove metadata, not behaviour. A pinned MCP tool's description and schema can't change unnoticed. Its code and answers can.
  • Project tokens are service identities. They don't prove which person acted. Named accounts do.
  • An approved call runs once per ticket, inside Guard. Your write tools still need idempotency keys: Guard can't guarantee exactly-once for an outside system.
  • The hosted control plane has no high availability. It runs as a single node. See Hosted control plane.
  • The scenarios in our demo workbench are fixed replays of recorded model output. They show the enforcement path working. They are not live attack tests against a model.

Reporting a security issue ​

Use our contact page. Don't include real secrets or personal data in your first message.