Reference

Security model

Default deny, user credentials, approvals, pinned host, limits, origins and redaction.

  • Default deny. Only allowlisted operations are exposed to the model.
  • User credentials. Your API sees the signed-in user's own credentials, never a service account.
  • Approvals. Non-GET calls need server-verified approval, bound to the exact arguments.
  • Pinned host. API calls go to baseUrl only. SSRF guards block private and metadata addresses, and redirects are not followed.
  • Limits. Size, step and time limits apply to every turn.
  • Origins. Browser calls from origins you did not list are refused.
  • Threads. Chat history is bound to the user and only shown to them.
  • Redaction. Audit events are redacted. LLM keys, approval secrets and user tokens do not appear in logs or errors.
  • Untrusted text. Knowledge passages and API results are passed to the model as data, and rendered as plain text, never HTML.